In progress – view active session
Conference 13407
Computer-Aided Diagnosis
16 - 20 February 2025 | Town & Country C
Conference Co-Sponsor
Show conference sponsors +
View Session 
SPIE Medical Imaging Awards and Plenary
16 February 2025 • 5:30 PM - 6:30 PM PST | Town & Country B/C
Session Chairs:
Joseph Y. Lo, Carl E. Ravin Advanced Imaging Labs. (United States), Cristian A. Linte, Rochester Institute of Technology (United States)
View Full Details: spie.org/medical-imaging-awards-plenary
5:30 PM - 5:40 PM:
Symposium Chair Welcome and Best Student Paper Award announcement
First-place winner and runner-up of the Robert F. Wagner All-Conference Best Student Paper Award
Sponsored by:
MIPS and SPIE
5:40 PM - 5:45 PM:
New SPIE Fellow acknowledgments
Each year, SPIE promotes Members as new Fellows of the Society. Join us as we recognize colleagues of the medical imaging community who have been selected.
SPIE Harrison H. Barrett Award in Medical Imaging
Presented in recognition of outstanding accomplishments in medical imaging
13408-500
Development and translation of 3D ultrasound-based imaging systems for diagnostic and image-guided interventions
(Plenary Presentation)
16 February 2025 • 5:50 PM - 6:30 PM PST | Town & Country B/C
Show Abstract +
Our research has been focused on developing 3D US scanning devices that overcome the limitations of conventional US imaging methods. We have been developing and fabricating various mechanical external motorized fixtures that move a conventional US probe in specific patterns and used them in systems for image-guided prostate biopsy prostate, prostate and gynecologic brachytherapy, and focal liver tumour ablation. As well, we developed 3D US-based system for point of care diagnostic application such as whole breast imaging, carotid plaque quantification, and hand and knee osteoarthritis.
Our approach allows scanning the desired anatomy in a consistent manner, imaging a large volume, integration of any manufacturer's 2D US probe into our fixtures, and integration of machine learning methods for rapid diagnosis and guidance. This approach provides a means of using US images with any US system with a small additional cost and minimal environmental constraints.
All-Symposium Welcome Reception
16 February 2025 • 6:30 PM - 8:00 PM PST | Flamingo Lawn
View Full Details: spie.org/mi/welcome-reception
Join your colleagues on the lawn for food and drinks as we welcome each other to SPIE Medical Imaging 2025.
Monday Morning Keynotes
17 February 2025 • 8:20 AM - 10:30 AM PST | Town & Country B/C
Session Chairs:
Ke Li, The Univ. of Texas M.D. Anderson Cancer Ctr. (United States), Mark A. Anastasio, Univ. of Illinois (United States), Shandong Wu, Univ. of Pittsburgh (United States)
View Full Details: spie.org/monday-morning-keynotes
8:20 AM - 8:25 AM:
Welcome and introduction
8:25 AM - 8:30 AM:
Award announcements
- Robert F. Wagner Award finalists for conferences 13405, 13409, and 13411
- Physics of Medical Imaging Student Paper Award
13405-501
Promoting innovation in your team: Lessons learned from 40 years in medical imaging
(Keynote Presentation)
17 February 2025 • 8:30 AM - 9:10 AM PST | Town & Country B/C
Show Abstract +
The development of advanced cross-sectional imaging technologies, especially X-ray CT and MRI, are widely recognized as the most impactful inventions in health care during the last 50 years. During this period of transformative innovation in medical imaging, progress has been accelerated through collaborative efforts between medical physicists, physicians, and the medical imaging industry. Innovation can be accelerated through individual efforts to promote the creative process, as well as frameworks to enhance collaboration and invention amongst teams of researchers. The purpose of this lecture is to examine key elements of the inventive process that have contributed to the development of medical imaging in the past that can be leveraged for ongoing advances in healthcare in the future.
13409-502
Designing AI for clinical imaging: the important role of model observers
(Keynote Presentation)
17 February 2025 • 9:10 AM - 9:50 AM PST | Town & Country B/C
Show Abstract +
Deep learning algorithms for image reconstruction and processing are showing strong promise for multiple medical-imaging applications. However, medical images are acquired for clinical tasks, such as defect detection and feature quantification, and these algorithms are often developed and evaluated agnostic to this clinical task. This talk will demonstrate how model observers can facilitate the development and evaluation of deep learning algorithms for clinical tasks by presenting two case studies. The first case study will underscore the misleading interpretations that clinical-task-agnostic evaluation of AI algorithms can yield, emphasizing the crucial need for clinical-task-based evaluation. Next, we will see how model observers can not only facilitate such evaluation but also enable the designing of deep learning algorithms that explicitly account for the clinical task, thus poising the algorithm for success in clinical applications. The second case study will demonstrate the use of model observers to select deep learning algorithms for subsequent human-observer evaluation. We will then see how this led to the successful evaluation of a candidate algorithm in a multi-reader multi-case human observer study. These case studies will illustrate how model observers provide a practical, reliable, interpretable, and efficient mechanism for development and translation of AI-based medical imaging solutions.
13411-503
Tackling the health AI paradox
(Keynote Presentation)
17 February 2025 • 9:50 AM - 10:30 AM PST | Town & Country B/C
Break
Coffee Break 10:30 AM - 11:00 AM
Session 1:
Classification
17 February 2025 • 11:00 AM - 12:20 PM PST | Town & Country C
Session Chairs:
Samuel G. Armato, The Univ. of Chicago (United States), Weijie Chen, U.S. Food and Drug Administration (United States)
13407-1
17 February 2025 • 11:00 AM - 11:20 AM PST | Town & Country C
Show Abstract +
In this paper, we propose a global-local hierarchical fusion method for automated GBC diagnosis with USG images. To overcome the issue of shallow features often being lost in deeper networks as the network depth increases, we designed the SAFM to fuse multi-scale features within a single branch. Since artifacts in USG images can cause structures to be inaccurately displayed, global information may fail to fully capture the target's details or contextual information. Therefore, we designed the HFM to fuse features from the same layer of the two branches, making the network more adaptable to complex tasks and scenarios. Finally, a fusion branch using transformer adaptively adjusts the importance of global and local features to improve the overall classification performance of the network. Comparative results against various DNN networks demonstrate that the network proposed in this paper outperforms all existing methods and enhances network stability.
13407-2
17 February 2025 • 11:20 AM - 11:40 AM PST | Town & Country C
Show Abstract +
Many medical diagnostic tasks have outcomes that are ordinal in nature and the features associated with these
ordinal outcomes are often also ordinal. This work seeks to develop methods to quantify the inherent ordinality in data in the interest of improving classification performance. We build on prior work that proposed summarizing feature ordinality with a pairwise separability matrix (PSM). We propose new distance measures to calculate the PSM and to summarize the apparent ordinality of the PSM into a single number reflective of the ordinality present in data. We validate these methods with 1) multiple variations of simulated data for a four-class ordinal task and with 2) radiomics features extracted from a multi-institutional set of N=660 chest x-ray radiographs acquired to identify four grades of pneumoconeosis. Results demonstrate that PSM’s can be reliably calculated based on separation of features in the original feature space using conventional distance measures. Likewise, we show that our proposed summary measure estimates overall ordinality that correlates well with known levels of ordinality in both the simulated and real data.
13407-3
17 February 2025 • 11:40 AM - 12:00 PM PST | Town & Country C
Show Abstract +
This study applied spherical harmonics-based convolutional neural network approach (SPHARM-Net) to MRI-derived brain shape metrics to predict age, sex, and Alzheimer’s disease (AD) diagnosis. MRI-derived brain features included vertex-wise cortical curvature, convexity, thickness, and surface area. SPHARM-Net performs convolutions using the spherical harmonic transforms, eliminating the need to explicitly define neighborhood size, and achieving rotational equivariance. Sex classification and age regression were carried out in a large sample of healthy adults (UK Biobank; N=32,979), and AD classification performance was tested in a large, publicly available sample (ADNI; N=1,213). SPHARM-Net showed strong performance for sex classification (accuracy=0.91; balanced accuracy= 0.91; AUC=0.97), and age regression (average absolute error=2.97 years; R-squared=0.77; Pearson's coefficient=0.9). AD classification also performed well (accuracy=0.86; balanced accuracy=0.83; AUC=0.9). Our experiments demonstrate promising preliminary performance using the SPHARM-Net for two widely studied benchmarking tasks and for AD classification.
13407-4
17 February 2025 • 12:00 PM - 12:20 PM PST | Town & Country C
Show Abstract +
This study addresses the high misdiagnosis rate of follicular thyroid carcinoma by developing a knowledge-driven deep learning framework (KDDL), which integrates a knowledge guide module (KGM) and follicular nodule capsule networks (FNCN), for achieving ultrasound thyroid nodule high classification accuracy. Using our innovative SRI-LDF (segmented ROI image fused with LBP and DWT) enhanced data as input, the output features of KGM can effectively complement FNCN and guide KDDL to achieve final classification prediction of follicular thyroid nodule images, thereby providing significant diagnostic support for FTC.
Break
Lunch Break 12:20 PM - 1:40 PM
Session 2:
Abdomen I
17 February 2025 • 1:40 PM - 3:00 PM PST | Town & Country C
Session Chairs:
Lubomir M. Hadjiiski, Michigan Medicine (United States), Amber L. Simpson, Queen's Univ. (Canada)
13407-5
17 February 2025 • 1:40 PM - 2:00 PM PST | Town & Country C
Show Abstract +
Liver cirrhosis, a leading cause of global mortality, requires precise segmentation of ROIs for effective disease monitoring and treatment planning. Existing segmentation models often fail to capture complex feature interactions and generalize across diverse datasets. To address these limitations, we propose a novel synergistic theory that leverages complementary latent spaces for enhanced feature interaction modeling. Our proposed architecture, \textit{nnSynergyNet3D}, integrates continuous and discrete latent spaces for 3D volumes and features auto-configured training. This approach captures both fine-grained and coarse features, enabling effective modeling of intricate feature interactions. We empirically validated \textit{nnSynergyNet3D} on a private dataset of 628 high-resolution T1 abdominal MRI scans from 339 patients. Our model outperformed the baseline \textit{nnUNet3D} by approximately 2\%. Additionally, zero-shot testing on healthy liver CT scans from the public LiTS dataset demonstrated superior cross-modal generalization capabilities. These results highlight the potential of synergistic latent space models to improve segmentation accuracy and robustness, thereby enhanc
13407-6
17 February 2025 • 2:00 PM - 2:20 PM PST | Town & Country C
Show Abstract +
This study proposes to advance the detection of gastric cancer by trianing YOLOv8 architecture through the generated synthetic gastric cancer images using the StyleGAN3 architecture and evaluating their effectiveness in AI-assisted diagnostics. To address this, the research utilizes styleGAN3 architecture based on generative adversarial networks to produce high-resolution synthetic images that enhance the training datasets for AI models. The study developed an image generative model that creates synthetic gastric cancer images at 1024x1024 resolution and used the pseudo-labeling method in a semi-supervised learning method to improve trained detection models. The generative model's image quality is validated through quality metrics of Fréchet Inception Distance, achieved a score of 34.71, it showed indicating high-quality synthesis image. By combining real and synthetic datasets, the researchers developed a gastric cancer detection model with a sensitivity of 90.30%, outperforming models trained solely on real (85.95%) or synthetic data (82.20%).
13407-7
17 February 2025 • 2:20 PM - 2:40 PM PST | Town & Country C
Show Abstract +
To improve the segmentation performance of kidney tumors in abdominal CT images, this study proposes two hybrid methods that combine convolutional neural networks (CNNs) with transformer-based network. Method A extracts features using a CNN block and then uses them as inputs to a transformer-based network. Method B employs a dual encoder structure with both CNN and transformer encoders. Experimental results showed that Method B significantly outperforms Method A in terms of balanced accuracy, Dice Similarity Coefficient (DSC). Method B demonstrates improved detection and segmentation of small tumors, reducing over-segmentation and substantially enhancing under-segmentation compared to Method A.
13407-8
17 February 2025 • 2:40 PM - 3:00 PM PST | Town & Country C
Show Abstract +
We are developing methods to estimate the ML/AI model output confidence in treatment response assessment for bladder cancer in CT urography. The model output confidence was estimated by an ensemble of ML/AI models, each incorporating a different lesion segmentation algorithm. The cases were then split into “easy” group with smaller variability and “difficult” group with larger variability in the outputs of the model ensemble. The AUC of the “difficult” cases was lower (AUC range: 0.58-0.80) compared to the AUC of the “easy” cases (AUC range: 0.85-0.92) for the radiomics model. The trend was consistent for the different methods of variability estimation. This indicates the feasibility of using the proposed methods for the estimation of model output confidence.
Break
Coffee Break 3:00 PM - 3:30 PM
Session 3:
Head, Neck, and Eye
17 February 2025 • 3:30 PM - 5:30 PM PST | Town & Country C
Session Chairs:
Thomas Martin Deserno, Peter L. Reichertz Institut für Medizinische Informatik (Germany), Leticia Rittner, Univ. of Campinas (Brazil)
13407-9
17 February 2025 • 3:30 PM - 3:50 PM PST | Town & Country C
Show Abstract +
Segmentation of thin structures such as the orbital medial wall and orbital floor poses challenges due to their low intensity and ambiguous boundaries. To address these issues, we propose a novel approach to segment these complex anatomical structures by using a diffusion model and refining the segmentation results with a consensus score and information extracted from CT images. We train the diffusion model using three annotation masks to account for inter-observer variability. To refine the segmentation results, we employ a consensus score map that indicates the agreement among 100 segmentation results, along with information from the CT images, including distance, brightness, and gradient directionality. Experimental results demonstrate that our method improves DSC, recall, and precision, and effectively mitigates issues such as discontinuities and overly thin segmentation that can occur from commonly used method that simply averages outputs of a diffusion model.
13407-10
Graph-based representation of retinal lesions for an interpretable diagnosis of diabetic retinopathy
17 February 2025 • 3:50 PM - 4:10 PM PST | Town & Country C
Show Abstract +
We present a novel graph-based approach for grading diabetic retinopathy (DR) that combines lesion segmentation and graph neural networks (GNNs). By representing lesions as nodes connected by k-nearest neighbors, our method aligns with clinical guidelines while leveraging deep learning. Evaluated on public datasets, our approach achieves performance comparable to state-of-the-art methods, offering potentially enhanced interpretability and robustness for clinical use.
13407-11
17 February 2025 • 4:10 PM - 4:30 PM PST | Town & Country C
Show Abstract +
Contrast-enhanced T1-weighted magnetic resonance imaging (MRI) with gadolinium-based contrast agents (GBCAs) offers superior discrimination between nasopharyngeal carcinoma (NPC) tissue and non-malignant tissue, making it crucial for NPC diagnosis. However, due to concerns about gadolinium accumulation, there is a global interest in developing contrast agent-free alternatives to replace GBCA-enhanced MRI. In this study, we developed an NPC diagnosis model using non-contrast images under the guidance of contrast-enhanced T1-weighted images. Our model was trained on a dataset of 694 cases and validated on 160 cases, with an equal split of 50% NPC and 50% non-NPC for both training and validation. It was then tested on an additional dataset comprising 263 cases (43% NPC and 57% non-NPC). The proposed model demonstrated high accuracy in detecting NPC using only non-contrast images, achieving performance comparable to predictions based on both contrast and non-contrast images.
13407-12
17 February 2025 • 4:30 PM - 4:50 PM PST | Town & Country C
Show Abstract +
Vision-influencing ocular conditions such as diabetic retinopathy or age-related macular degeneration affect millions globally, and early diagnosis through color fundus photographs (CFPs) can mitigate their impact. However, the manual interpretation of CFPs is time-consuming and requires skilled ophthalmologists. Recent discriminative deep learning models have achieved high accuracy in automatically detecting ocular conditions from CPFs, but they are often seen as black boxes that are incapable of explaining their decisions convincingly. To address this, we propose a generative conditional diffusion model that detects ocular conditions in CFPs and generates counterfactual images to visualize its decision-making process. We show that this novel self-explainable generative classifier achieves competitive classification accuracy while the synthesized counterfactuals confirm that it has learned hallmark features of the ocular conditions tested. We believe that our approach is an important step towards trustworthy AI in ophthalmology as it removes the explainability challenges surrounding standard classifiers.
13407-13
17 February 2025 • 4:50 PM - 5:10 PM PST | Town & Country C
Show Abstract +
This study presents IMLSP, an Interpretable Multi-Label multi-modal deep Survival Prediction framework to enhance personalized management for Head and Neck Cancer (HNC) patients undergoing Radiation Therapy (RT). IMLSP predicts multiple survival outcomes simultaneously while providing visual explanations. The framework uses Multi-Task Logistic Regression (MTLR) layers to convert survival prediction into a multi-time-point classification task, enabling several relevant outcome predictions. The study introduces Grad-Team, a Gradient-weighted Time-event activation mapping technique for deep survival model visualization, generating patient-specific time-to-event activation maps. The method was evaluated using the RADCURE HNC dataset, where it outperformed single-modal and single-label models. Activation maps showed the model focused on tumor and nodal volumes, varying between high- and low-risk patients. Our findings indicate that multi-label learning enhances prognostic performance, and the visualization technique provides insights into AI decision-making, fostering trust in personalized cancer management.
13407-14
17 February 2025 • 5:10 PM - 5:30 PM PST | Town & Country C
Show Abstract +
Visual acuity is an important ophthalmologic measure. Its standard assessment method relies on vision tests, while there is no clinical method to derive visual acuity from medical eye images. This can be explained by the lack of defined structure-function correlations for all biomarkers visible in these images. Prior works showed that deep learning methods allow the prediction of visual impairment from medical images without biomarker identification. Beyond that, we show that fine-tuning an ophthalmic foundation model with a comparatively small data set from clinical routine enables us to derive visual acuity from only a single image. We adapt the foundation model RETFound such that it outputs one of three visual impairment levels from optical coherence tomography images taken of patients with one of two macular diseases. In this way, we achieve a satisfactory visual acuity prediction based on a single image and requiring only a small set of fine-tuning data.
Tuesday Morning Keynotes
18 February 2025 • 8:30 AM - 10:00 AM PST | Town & Country B/C
Session Chairs:
Jhimli Mitra, GE Research (United States), Christian Boehm, ETH Zurich (Switzerland)
View Full Details: spie.org/tuesday-morning-keynotes
8:30 AM - 8:35 AM:
Welcome and introduction
8:35 AM - 8:40 AM:
Award announcements
- Robert F. Wagner Award finalists for conferences 13406 and 13412
- Image Processing Student Paper Award
13406-504
Integrating computational approaches to unravel alzheimer's disease and co-pathologies: a biomarker-driven approach to precision medicine
(Keynote Presentation)
18 February 2025 • 8:40 AM - 9:20 AM PST | Town & Country B/C
Show Abstract +
Early detection and intervention in neurodegenerative diseases hold the potential to significantly impact patient outcomes. This presentation will explore the development of multi-disciplinary and multi-modality biomarkers to identify individuals at risk and monitor disease progression. By combining advanced imaging techniques, such as MRI, and PET, with fluid biomarkers, we aim to detect subtle changes in brain structure and function that precede clinical symptoms. These biomarkers could serve as powerful tools for early diagnosis, enabling timely intervention and potentially delaying disease onset. Furthermore, by identifying individuals at highest risk, we can optimize the design of clinical trials and accelerate the development of effective therapies. Ultimately, our goal is to improve the lives of individuals with neurodegenerative diseases through early detection, precise diagnosis, and targeted treatment.
13412-505
Wearable ultrasound technology
(Keynote Presentation)
18 February 2025 • 9:20 AM - 10:00 AM PST | Town & Country B/C
Show Abstract +
The use of wearable electronic devices that can acquire vital signs from the human body noninvasively and continuously is a significant trend for healthcare. The combination of materials design and advanced microfabrication techniques enables the integration of various components and devices onto a wearable platform, resulting in functional systems with minimal limitations on the human body. Physiological signals from deep tissues are particularly valuable as they have a stronger and faster correlation with the internal events within the body compared to signals obtained from the surface of the skin. In this presentation, I will demonstrate a soft ultrasonic technology that can noninvasively and continuously acquire dynamic information about deep tissues and central organs. I will also showcase examples of this technology's use in recording blood pressure and flow waveforms in central vessels, monitoring cardiac chamber activities, and measuring core body temperatures. The soft ultrasonic technology presented represents a platform with vast potential for applications in consumer electronics, defense medicine, and clinical practices.
Break
Coffee Break 10:00 AM - 10:30 AM
Session 4:
CAD and Perception: Joint Session with Conferences 13407 and 13409
18 February 2025 • 10:30 AM - 12:40 PM PST | Town & Country C
Session Chair:
Susan M. Astley, The Univ. of Manchester (United Kingdom)
13409-16
Does concurrent reading with AI lead to more false negative errors for cancers that are not marked by AI?
(Invited Paper)
18 February 2025 • 10:30 AM - 11:00 AM PST | Town & Country C
Show Abstract +
The purpose was to determine if reading digital breast tomosynthesis (DBT) concurrently with an artificial intelligence (AI) system increases the probability of missing a cancer not marked by AI for cancers that the radiologist detected reading without AI. We retrospectively analyzed a dataset from an observer study where 24 radiologists read 260 DBT screening exams (65 exams with cancer), with and without an AI system. We examined only cases that the radiologist recalled when reading without AI and grouped them by AI-detected and AI-missed, separately for cancer and non-cancer cases. When reading with AI concurrently, readers found 3.3 (46%) of the 7 AI-missed cancers and agreed with 54.2 (93%) of the 58 AI-detected cancers. Using a two-tailed, paired t-test, this difference (46% vs 93%) was statistically significant (p<<0.00001). Similarly, for non-cancer cases: if AI did not mark an abnormality in the image, radiologists were more likely to call the case normal, even though they called it abnormal when reading without AI (16% vs 65%, p<<0.00001). This explained nearly all the increase in specificity when reading in concurrent mode.
13407-15
18 February 2025 • 11:00 AM - 11:20 AM PST | Town & Country C
Show Abstract +
The accurate staging of esophageal cancer is critical for effective treatment planning. This study introduces an unsupervised methodology for classifying TNM categories in PET scans by using gradient-weighted class activation mapping (Grad-CAM) and uniform manifold approximation and projection (UMAP) with deep metabolic texture analysis. Using a patch-based Convolutional Neural Network (CNN) pre-trained on esophageal primary tumor (T) data, we applied Grad-CAM to identify significant regions in PET scans, followed by UMAP for dimensionality reduction. KMeans clustering was then utilized to classify the reduced embeddings into TNM categories. Our unsupervised approach addresses the challenge of limited annotated datasets available for nodes (N) and metastasis (M) detection by eliminating the need for extensive labeled datasets required for supervised learning. Our method demonstrated an accuracy and F1 score of 89.5% and 93.1%, respectively, in differentiating between primary tumors, nodes, and metastases. The results indicate significant potential for AI-led staging and personalized treatment planning.
13409-17
18 February 2025 • 11:20 AM - 11:40 AM PST | Town & Country C
Show Abstract +
This study examined the predictive power and causal contribution of calcification features in the deep-learning based Mirai breast cancer risk prediction model. To do this, we constructed "CalcMirai", a reduced version of that focusses solely on features related to calcifications. The CalcMirai model was used to conduct a selective mirroring experiment that considers only one breast, either the future cancerous breast or the healthy side of a patient with confirmed cancer, to predict the patient’s breast cancer risk. Our results showed that both Mirai and CalcMirai performed similarly well on breasts in which cancer will develop in the future. Mirroring the healthy breast reduced predicted risk for both models to a similar extent. The performance remained discriminative overall. This suggests that the predictive power of Mirai largely stems from the detection of early micro-calcifications and/or identifying high-risk calcifications.
13407-16
18 February 2025 • 11:40 AM - 12:00 PM PST | Town & Country C
Show Abstract +
Developing accurate machine learning methods for predicting breast cancer risk is reliant on the availability of good quality datasets on mammograms. For density prediction, these datasets require two experts to give their opinions about the VAS score that each example should receive. In many of these examples, the two experts disagree, sometimes quite substantially on what the correct VAS score should be. It has been found that by filtering the dataset based on the disagreement between experts can lead to a slight increase in the accuracy of the model when the predicted VAS score is used for computing the risk.
13409-18
18 February 2025 • 12:00 PM - 12:20 PM PST | Town & Country C
Show Abstract +
A newly proposed artificial intelligence (AI)-based tool, Smart Imagery Framing and Truthing (SIFT), was applied to provide lesion annotation of pulmonary abnormalities (or diseases) and their corresponding boundaries on 452,602 chest X-ray (CXR) images from four publicly available datasets. SIFT is based on Multi-task, Optimal-recommendation, and Max-predictive Classification and Segmentation (MOM ClaSeg) technologies to identify and delineate 65 different abnormalities. The MOM ClaSeg System is developed on a training dataset of over 300,000 CXR images, which contains over 240,000 confirmed abnormal images with over 300,000 confirmed ROIs corresponding to 65 different abnormalities and over 67,000 normal (i.e., “no finding”) images. SIFT system can determine the abnormality types of labeled ROIs and their boundary coordinates with high efficiency (improved 5.88 times) when radiologists used SIFT as an aide compared to radiologists using a traditional semi-automatic method. The SIFT system achieves an average sensitivity of 89.38%±11.46% across four datasets. This can be used to significantly improve the quality and quantity of training and testing sets to develop AI technologies.
13407-17
18 February 2025 • 12:20 PM - 12:40 PM PST | Town & Country C
Show Abstract +
This study developed an automated method for segmenting microvascular density regions in OCT-angiography (OCTA) images using deep learning. Four models with different input combinations were compared to determine if additional inputs improved prediction accuracy. The dataset included 50 training and 47 test images labeled by two experts. Results showed no significant differences between Expert 1 and the models, but visual inspection suggested that the model with three-channel input (OCTA + foveal avascular zone + large vessel tree) occasionally produced more consistent results. ANOVA tests compared the Dice coefficients for Expert 1, Expert 2, and the three-channel input model and found significant differences only in the normal category (p-value: 0.036), while Tukey’s HSD test showed no significant differences between each comparison. This automated approach offers a reliable alternative to manual assessments, providing consistent and objective measurements for capillary density in OCTA images.
Break
Lunch Break 12:40 PM - 1:50 PM
Session 5:
Cardiovascular
18 February 2025 • 1:50 PM - 3:10 PM PST | Town & Country C
Session Chairs:
Khan M. Iftekharuddin, Old Dominion Univ. (United States), Chisako Muramatsu, Shiga Univ. (Japan)
13407-18
18 February 2025 • 1:50 PM - 2:10 PM PST | Town & Country C
Show Abstract +
Atherosclerosis is a hardening process of the arteries caused by plaque buildup in and on the artery wall. The risk of cardiovascular diseases is correlated with the extent of atherosclerotic plaque burden. This work targets automatic calcified plaque detection and segmentation to assess the plaque burden. It is formulated as multi-class segmentation of plaque, aorta, body, and background on CT scans using the state-of-the-art deep learning framework, nnUNet. Agatston scores calculated on the segmented plaque regions are used to assess the plaque burden. Experimental results on both non-contrast and contrast enhanced CT scans showed that calcified plaques were accurately detected and segmented. The plaque burden was also accurately assessed, potentially leading to improved atherosclerotic disease diagnosis.
13407-19
18 February 2025 • 2:10 PM - 2:30 PM PST | Town & Country C
Show Abstract +
In cardiac CT angiography imaging, extracting coronary centerlines can be a tedious and time-consuming task for radiologists. In this paper we propose a DL method which is able to automatically track vessels from seed points to their ends without manual intervention. Our network consists of two lightweight 3D convolutional neural networks (CNNs), one trained to predict forward and backward directions and the other to generate a vessel distance map (DMAP). The DMAP-CNN is important for vessel tracking, which not only aids in extracting seed points but also provides a robust stopping mechanism before the tracking extends beyond the vessel of interest. The proposed networks were trained using 55 patient data, 2 for validation data and 5 for inference. Our experimental studies show that the overall performance of the proposed method for the RCA and LCA has a misclassification rate of about 14%, with mean sensitivity and overlap of 95%. These results demonstrate that our method can reliably track coronary arteries.
13407-20
18 February 2025 • 2:30 PM - 2:50 PM PST | Town & Country C
Show Abstract +
Deep learning (DL)-based segmentation of cerebral vessels is gaining popularity. Most DL based methodologies are only trained and tested on a single-center dataset; however, for clinical translation, generalizable methods are required. We present a human-in-the-loop transfer-learning approach to address this limitation. Using a DL model, pre-trained on an internal cohort of computed tomography angiography (CTA) images (n=50), we observed how a cohort of external images (n=3 to 9 CTA) can help tune the model to accurately segment vessels on an external CTA dataset. This has a direct impact on the clinical applicability of DL models for cerebral vessel segmentation.
13407-21
18 February 2025 • 2:50 PM - 3:10 PM PST | Town & Country C
Show Abstract +
AI models with high performance metrics often falter in clinical practice due to confounding factors like imbalanced data, variable image quality, and inconsistent clinical labels. Here, we explored outlier detection in priorly identifying patients with clinical confounders. An example uses coronary angiography to monitor reduced left ventricular ejection fraction (LVEF). We recently developed an ensemble model of 3D-CNN and Transformer, to distinguish patients with reduced LVEF (≤40%) from normal LVEF (>40%). Despite achieving an AUC of 0.87, the model’s accuracy was 77%, lower than imbalance ratio (82%). To address this, we employed outlier detection using KNN-cosine similarity of feature layers, correlating these scores with prediction accuracy. A reader study confirmed that identified inliers had superior image quality and label consistency compared to outliers. We further applied conformal prediction to categorize predictions into ‘certain’ and ‘uncertain’ groups. Accuracy improved to 91% for the ‘certain’ group, while ‘uncertain’ cases showed reduced accuracy.
Break
Coffee Break 3:10 PM - 3:50 PM
Session 6:
Chest
18 February 2025 • 3:50 PM - 5:30 PM PST | Town & Country C
Session Chairs:
Maryellen L. Giger, The Univ. of Chicago (United States), Horst Karl Hahn, Fraunhofer-Institut für Digitale Medizin MEVIS (Germany), Jacobs Univ. Bremen (Germany)
13407-22
18 February 2025 • 3:50 PM - 4:10 PM PST | Town & Country C
Show Abstract +
Advancements in artificial intelligence (AI) have improved the efficiency and accuracy of radiological examinations, but creating accurate AI models is still challenging due to the requirement for extensive manual annotation. We propose the Smart Imagery Framing and Truthing (SIFT) system to assist annotators in labeling and delineating pulmonary abnormalities on CT images. SIFT is developed on a training dataset of 9,078 CT images (18,367 ROIs) corresponding to 47 different abnormalities/diseases. An independent testing set with 2,199 CT images (4,280 ROIs) is processed by SIFT to predict both abnormality/disease types and ROI boundary locations. Evaluation metrics include IOU, AUC, and slice-level sensitivity of abnormality labels. For ROI segmentation, 91.5% and 36.2% of abnormalities had an IOU over 0.6 and 0.7, respectively. Regarding AUC, 97.9%, 80.9%, and 42.6% of abnormalities had values above 0.7, 0.8, and 0.9, respectively. For sensitivity, all 47 label categories exceeded 0.8. SIFT demonstrates high performance in determining abnormality/disease types with corresponding boundary locations for the ROIs, which can be used to predict training and testing sets to develop AI.
13407-23
18 February 2025 • 4:10 PM - 4:30 PM PST | Town & Country C
Show Abstract +
The increasing number of CT scan examinations coupled with the time-intensive nature of manual analysis highlights the need for efficient automated analysis methods to assist radiologists in detecting anomalies. While state-of-the-art deep learning approaches directly classifies abnormalities from 3D CT images, radiologists also rely on clinical indications and patient demographics such as age and sex for diagnosis. We propose a multimodal model that integrates 3D chest CT scans, clinical information reports, patient age and sex to improve multi-label abnormality classification. Our approach leverages both imaging and non-imaging data by combining visual and textual networks. We extend our work by conducting an ablation study on a public dataset to evaluate and demonstrate the efficacy of our approach.
13407-24
18 February 2025 • 4:30 PM - 4:50 PM PST | Town & Country C
Show Abstract +
Chest X-ray (CXR) imaging is crucial for identifying pathological changes in the thoracic cavity. AI and ML applications rely on accurate detection of anatomical structures in CXRs for screening/diagnostic applications. The YOLO object detection models have recently gained prominence for detecting anatomical structures in medical images, however the state-of-the-art results are limited to single-organ detection. In this work we propose a multi-organ detection technique through two recent YOLO versions and their sub-variants. We evaluate their effectiveness in detecting lung and heart regions in CXRs simultaneously. For this we used the JSRT CXR dataset for internal validation and other data sets for external testing. We find that YOLOv9 models outperform YOLOv8 variants, with ensemble approaches further improving detection.
13407-25
18 February 2025 • 4:50 PM - 5:10 PM PST | Town & Country C
Show Abstract +
This study examines the use of transformer-based networks with declarative knowledge in assessing endotracheal tube (ETT) placement in chest radiographs. The optimal placement of the ETT is crucial for patient safety, and it typically resides 5±2 cm above the Carina. Misplacement of ETT can lead to severe complications, heightening the need for accurate and timely verification methods in intensive care settings. In this work, we propose an AI system with Transformer-based UNet with declarative knowledge trained and evaluated on a dataset of 200 anonymized chest X-ray images to determine the ETT position effectively. These networks localize the tip of the ETT, and Carina, ensuring the tube’s placement within a predefined safe zone. Expanding on the limitations previous works in literature, this proposed system is more robust and less prone to error. The initial findings are encouraging and motivate us to work further to be able to effectively integrate the proposed new system into clinical practice for real-time use.
13407-26
18 February 2025 • 5:10 PM - 5:30 PM PST | Town & Country C
Show Abstract +
We propose a novel curriculum learning (CL) approach for lung nodule diagnostic interpretation. Our CL approach embeds an easy-to-moderate-to-hard case strategy into the incremental learning process by starting the learning process using In-Distribution (ID) cases and then adds iteratively easy to moderate to hard cases determined based on their Out-of-Distribution (OoD) score. Using the NIH/NCI Lung Image Database Consortium (LIDC) data, we show that the CL approach improves the classification of the minority class (malignant cases) even for the out-of-distribution (OoD) samples as well as when the inter-observer interpretation variability is high. These results are significant because they show the potential of curriculum learning for improving the performance of Computer-aided Diagnosis (CAD) systems in the presence of unbalanced datasets and class label uncertainty.
Live Demonstrations Workshop
18 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View Full Details: spie.org/live-demonstrations-workshop
The goal of this workshop is to provide a forum for systems and algorithms developers to show off their creations. The intent is for the audience to be inspired to conduct derivative research, for the demonstrators to receive feedback and find new collaborators, and for all to learn about the rapidly evolving field of medical imaging. The Live Demonstrations Workshop invites participation from all attendees of the SPIE Medical Imaging symposium. Workshop demonstrations include samples, systems, and software demonstrations that depict the implementation, operation, and utility of cutting-edge as well as mature research. Having an accepted SPIE Medical Imaging paper is not required for giving a live demonstration. A certificate of merit and $500 award will be presented to one demonstration considered to be of exceptional interest.
Award sponsored by:
Siemens Healthineers
Wednesday Morning Keynotes
19 February 2025 • 8:30 AM - 10:00 AM PST | Town & Country B/C
Session Chairs:
Maryam E. Rettmann, Mayo Clinic (United States), Aaron D. Ward, The Univ. of Western Ontario (Canada)
View Full Details: spie.org/wednesday-morning-keynotes
8:30 AM - 8:35 AM:
Welcome and introduction
8:35 AM - 8:40 AM:
Award announcements
- Robert F. Wagner Award finalists for conferences 13408 and 13413
- Early-Career Investigator Award: Image-Guided Procedures, Robotic Interventions, and Modeling
- Student Paper Award: Image-Guided Procedures, Robotic Interventions, and Modeling
13408-506
Image-guided surgery and examinations using ultrasound
(Keynote Presentation)
19 February 2025 • 8:40 AM - 9:20 AM PST | Town & Country B/C
Show Abstract +
Many of today’s cancer surgeries are carried out with robot assistance. Using real-time intra-operative ultrasound, we can overlay pre-operative imaging into the surgeon’s console, enabling visualization of sub-surface anatomy and cancer at the same time with the standard laparoscopic camera view. We will discuss aspects of system design, visualization and registration methods that enable such visualization, and present our results. We will also present tissue and instrument tracking approaches that can be used in future augmented reality systems.
For remote and underserved communities, we developed a teleultrasound approach that relies upon using a novice – the patient, a family member or friend – as a robot to carry out the examination. The novice wears a mixed reality headset and follows a rendered virtual ultrasound transducer with the actual transducer. The virtual transducer is controlled by an expert, who sees the remote ultrasound images and feels the transducer forces. This tightly-coupled expert-novice approach has advantages relative to both conventional and robotic teleultrasound. We discuss our system implementation and results.
13413-507
The future of diagnostics: the role of computational pathology in tomorrow's medicine
(Keynote Presentation)
19 February 2025 • 9:20 AM - 10:00 AM PST | Town & Country B/C
Show Abstract +
Computational Pathology has already led to remarkable innovations in diagnostics, achieving expert pathologist performance in tasks such as prostate cancer grading and cancer metastasis detection. In recent years, we have seen rapid advances, with weakly supervised models able to predict patient outcomes or genetic mutations and foundation models enabling application to rarer diseases. However, this only scratches the surface of what will be possible in the near future.
In this talk, I will briefly touch on the history of computational pathology and how we got to where we are today. Subsequently, I will highlight the current methodological innovations in the field and their potential for causing a paradigm shift in diagnostic pathology. I will discuss how these innovations, combined with the AI-driven integration of radiology, pathology, and 'omics data streams, could change the future of diagnostics as a whole. Last, I will discuss the challenges and pitfalls moving forward and how we, as a community, can contribute to addressing them.
Break
Coffee Break 10:00 AM - 10:30 AM
Session 7:
Breast
19 February 2025 • 10:30 AM - 12:30 PM PST | Town & Country C
Session Chairs:
Juhun Lee, Univ. of Pittsburgh (United States), Heather M. Whitney, The Univ. of Chicago (United States)
13407-27
19 February 2025 • 10:30 AM - 10:50 AM PST | Town & Country C
Show Abstract +
Uncertainty estimations have the potential to increase the trustworthiness of artificial intelligence (AI) outputs. However, the interpretation of such estimates varies with the AI model and estimation approach used. In this work, we demonstrate the variability of four commonly used uncertainty estimation approaches to changes in model performance and model knowledge. We show that these four off-the-shelf methods incorporate different amounts of model uncertainty (lack of model knowledge) and data uncertainty (noise or ambiguity of the input) when estimating the predictive uncertainty. We conclude that the interpretation of AI uncertainty varies based on model performance and estimation method used.
13407-28
19 February 2025 • 10:50 AM - 11:10 AM PST | Town & Country C
Show Abstract +
This study introduces LLaVA-MultiMammo, a novel approach adapting vision-language models for comprehensive and explainable multi-view mammogram analysis. Unlike traditional computer-aided detection schemes, LLaVA-MultiMammo integrates multi-view imaging data with clinical information to perform multiple tasks within a single framework. Evaluated on BI-RADS density categorization, malignancy classification, and 5-year breast cancer risk prediction using the EMBED AWS Open Data, the model demonstrates competitive performance compared to task-specific approaches. LLaVA-MultiMammo provides natural language interactions, enhancing interpretability. This versatile tool could significantly improve how radiologists interact with AI-assisted diagnostic tools, advancing breast cancer screening and diagnosis.
13407-29
19 February 2025 • 11:10 AM - 11:30 AM PST | Town & Country C
Show Abstract +
Bias in AI training data can create bias in the model produced, this is of particular importance in medical
scenarios to ensure fair and equitable representation. This work explores the data cleaning practices used and
their effect on a model trained to predict breast density, a strong indicator of cancer risk. This work’s primary
focus is the removal of patient mammograms with more than the standard four views. This can be because the
breast was too large to fit in a single image, or because of a technical repeat. The results of this work show
that inclusion of this data does not hinder AI training, and their data alone is sufficient to train AI. Exploring
exclusion criteria is important as any systematic practice can leave particular demographics unrepresented in the
data. This is an important step towards fair representation in AI and ensuring its benefits can be seen by all.
13407-30
19 February 2025 • 11:30 AM - 11:50 AM PST | Town & Country C
Show Abstract +
This study addresses the critical issue of racial bias in AI models within medical imaging by investigating anatomical differences in mammographic images between Asian and White women. Using the EMBED dataset for model development and internal testing and the dataset from the University of Pittsburgh Medical Center for independent testing, we developed a Vision Transformer (ViT) model to predict race based on breast anatomical features on mammographic images, achieving AUCs of 0.80 and 0.71, respectively. Our findings reveal notable differences in breast size and moderate differences in breast shape between these groups. This study provides important insights into mitigating racial bias between Asian and White populations, underscoring the need for equitable AI performance across diverse clinical settings.
13407-31
19 February 2025 • 11:50 AM - 12:10 PM PST | Town & Country C
Show Abstract +
Medical out-of-distribution (OOD) detection poses great challenges due to the heterogeneity and unknown characteristics of medical data. Therefore, to train a novelty detector with only ID data available, learning high-quality “normality” features is the fundamental step to identify the OOD samples during inference. Unfortunately, it still lacks an effective way to identify the difference for several datasets from the same medical domain. The main challenge
lies in the inaccessibility to external medical datasets given the privacy concerns around sharing personally
identifiable information. Therefore, an efficient way of external dataset curation by identification of anomaly
without sharing data is desired, particularly for AI model generalization. Our proposed novel HAND architecture,
composed of CNN+transformer, discriminator, and gradient reversal, can be trained in an unsupervised
way to detect both intra- and inter-class anomalies from mammogram screening exams, with the AUC score of 0.94 and 0.85 for internal and external evaluation, achieves the best performance among VAE-based, GAN-based, and CNN+transformer baselines.
13407-32
19 February 2025 • 12:10 PM - 12:30 PM PST | Town & Country C
Show Abstract +
Existing studies on segmenting dense tissue in Digital Breast Tomosynthesis (DBT) remain limited, largely due to the challenges posed by the complexity of multi-slice variations, blurring, and out-of-plane artifacts. This study introduced a fully automated dense tissue segmentation algorithm in DBT image using a fully convolutional network that can also be used to segment fatty tissue and breast area as well. We employed 20 DBT scans from 20 normal patients (BIRADS 1) from the Breast-Cancer-Screening-DBT dataset. For establishing ground truth, one radiologist segmented breast dense tissue, and breast area mask in every slice of each DBT volume. We preprocessed each DBT volume slice-by-slice. Finally, we constructed 3-channel RGB images (ground truth images) by assigning breast area, fatty and dense area into R, G, B channels. Using the DBT images and ground truth, we fine‐tuned the SegNet pretrained for segmenting breast density from 2D mammograms to segment both the breast and the fibroglandular areas in digital breast tomosynthesis (DBT) images. Using a test set, we achieved a dice score of 0.84±0.08. In the DBT, our model performed better than the other models.
Break
Lunch Break 12:30 PM - 1:50 PM
Session 8:
Methods
19 February 2025 • 1:50 PM - 3:10 PM PST | Town & Country C
Session Chairs:
Kenji Suzuki, Institute of Science Tokyo (Japan), Ronald M. Summers, National Institutes of Health Clinical Ctr. (United States)
13407-33
19 February 2025 • 1:50 PM - 2:10 PM PST | Town & Country C
Show Abstract +
Roughly half of all patients with colorectal cancer will develop liver metastases and face a 3-year survival rate of less than 10%. Models trained on quantitative information extracted from contrast-enhanced computed tomography (CECT) have the potential to predict patients at high risk for recurrence, although variations in image acquisition, reconstruction, and contrast timing affect their generalizability. 135 patients from Memorial Sloan Kettering Cancer Center (n = 68) and MD Anderson Cancer Center (n = 67) were prospectively enrolled and underwent an additional phase CECT scan within ± 15 seconds of the routine portal venous phase using a controlled protocol, with systematic variations in scan timing, image acquisition, and image reconstruction. Reproducibility of radiomic features extracted from the liver parenchyma and largest metastasis was measured using Lin's CCC. An increase in scan delay magnitude, certain combinations of reconstruction parameters, and the extraction of features from the liver parenchyma, rather than the metastasis, decreased reproducibility across contrast timing.
13407-34
19 February 2025 • 2:10 PM - 2:30 PM PST | Town & Country C
Show Abstract +
We present an analysis of the performance and feature selection bias associated with several variants of K-best and minimum redundancy, maximum relevance feature selection methods when applied in a survival analysis setting. A Cox proportional hazards model was trained to predict overall survival based on radiomic features from 197 patients who underwent hepatic resection to treat colorectal liver metastases. We also considered the removal of features of low univariate significance, and features with low reproducibility in an independent data set. We found that using mRMR was associated with a reduction in performance C-index of 0.016 compared to K-best. Though mRMR models had lower bias, the highest bias was obtained when combining mRMR and univariate significance thresholding. Counterintuitively, we found that the simplest method, which ignored feature redundancy, and right-censoring, and implemented no pre-processing, resulted in the best performance (C-index=0.600 (0.591--0.607)), with low feature selection bias (0.000 (-0.001--0.004)).
13407-35
19 February 2025 • 2:30 PM - 2:50 PM PST | Town & Country C
Show Abstract +
Multi-task deep learning has been developed to ensure explainability in computer-aided diagnosis (CAD) systems.
These systems simultaneously predict interpretable characteristics of findings in auxiliary tasks to help radiologists assess the reliability of the diagnostic prediction.
However, the CAD's opinion would not be referred to when the auxiliary task prediction deviates from the radiologist's expectation.
To further leverage CAD's knowledge in the above situation, we proposed a multi-task gradient assist (MGA) method that brings interactivity to multi-task models without additional training.
MGA allows radiologists to give feedback to the CAD and ask for additional diagnostic suggestions.
Using two datasets of mammography and CT, we show that MGA successfully improves diagnostic accuracy upon feedback on various auxiliary tasks.
Compared to an existing method using forward propagation only, MGA performs more robustly and can be combined to further improve accuracy.
Our method paves a promising way for integrating artificial intelligence into image diagnosis.
13407-36
19 February 2025 • 2:50 PM - 3:10 PM PST | Town & Country C
Show Abstract +
AI models in healthcare often exhibit performance biases between demographic subgroups. These disparities can
originate from improper data collection, model training, and/or clinical deployment. A promising method to tackle
such biases lies reducing differences in training and test datasets. Here, we examine the effects of using stratified
sampling across demographic (age, race, ethnicity, and sex) and disease attributes to yield similar training and test
sets. To quantify similarity between the resultant subsets, we make use of the Jensen-Shannon distance (JSD) calculated for each attribute. We also use a multidimensional JSD, calculated using an aggregate method, to
quantify similarity across the combination of all attributes.
Break
Coffee Break 3:10 PM - 3:50 PM
Session 9:
Segmentation
19 February 2025 • 3:50 PM - 5:30 PM PST | Town & Country C
Session Chairs:
Shandong Wu, Univ. of Pittsburgh (United States), Karen Drukker, The Univ. of Chicago (United States)
13407-37
19 February 2025 • 3:50 PM - 4:10 PM PST | Town & Country C
Show Abstract +
Recent studies using DCNN based on U-Net for automatic muscle recognition have been limited to conventional anatomical planes. We propose a novel dual-perspective virtual unfolding U-Net model (DPVU-Net) utilizing virtual unfolding CT images. Our method creates images from anterior and posterior central body positions, allowing comprehensive body observation. We evaluated our approach on 30 non-contrast body CT cases, focusing on 3D segmentation of six muscle regions. The DPVU-Net achieved an average Dice score of 77.2 ± 8.5% across these regions, outperforming single-perspective methods (73.0 ± 9.9% and 73.8 ± 9.3%). By integrating two perspectives with different characteristics, our method demonstrated effectiveness in superficial muscle recognition. This study contributes to advancing CAD systems in CT images through a DCNN using virtual unfolding CT images as input.
13407-38
19 February 2025 • 4:10 PM - 4:30 PM PST | Town & Country C
Show Abstract +
In this work, we demonstrate that previous instance-based evaluation schemes such as Lesion-wise Dice from the Brain Tumor Segmentation (BraTS) team and the first instance based metric Panoptic Quality can produce both inflated or deflated instance based scores during the evaluation of instance imbalance in semantic segmentation. To address this limitation, we propose one of the first many-to-many matching schemes called Counter-dice. .Counter Dice removes the requirement of creating individual instances using connected component analysis. Instead, it uses connected component analysis of an overlay of predicted and ground truth masks and calculates the final score from the mean of all the created clusters based on the overlay. This simple framework is easy to define and implement, transparent, highly efficient to allow for rapid evaluation and can accommodate boundary-based metrics as well.
13407-39
19 February 2025 • 4:30 PM - 4:50 PM PST | Town & Country C
Show Abstract +
Chronic lower back pain (cLBP) research often neglects a comprehensive layer-by-layer analysis of anatomical structures. Annotating hundreds of slices in a 3D medical imaging examination for machine learning is time-consuming. To address this, we propose InterSliceBoost, a novel method comprising an inter-slice generator and a semantic segmentation model, designed to generate accurate layer-by-layer masks with minimal manual annotation. Our experiments demonstrated that the segmentation model using InterSliceBoost achieved comparable performance to full annotation when trained on only 33% of the B-mode ultrasound images.
13407-40
19 February 2025 • 4:50 PM - 5:10 PM PST | Town & Country C
Show Abstract +
Measuring skin involvement in chronic graft-versus-host disease (cGVHD) currently requires expert manual assessment, which is time-consuming and shows high interrater disagreement (>20% surface area). Automated image analysis has shown promise for measuring affected skin area under controlled photography conditions. We improve the performance in standard clinical photographs without additional expert annotations using a semi-supervised approach. A baseline U-Net model was trained using 3D photos from 36 cGVHD patients with expert demarcations. The model was then retrained with 2D photos from 83 additional patients using a semi-supervised method. Testing on 20 held-out patients, the median surface area error improved from 19.2% (interquartile range 6.3 – 33.8) at baseline to 10.2% (4.5 – 22.6) after retraining. This approach was effective for translating the pre-trained U-Net model to standard clinical photos without additional expert annotations. Such models could help standardize cGVHD assessment and alleviate expert burden.
13407-41
19 February 2025 • 5:10 PM - 5:30 PM PST | Town & Country C
Show Abstract +
The segmentation of multiple organs in multi-parametric MRI studies is critical for many applications in radiology, such as correlating imaging biomarkers with disease status (e.g., cirrhosis, diabetes). Recently, three publicly available tools, such as MRSegmentator (MRSeg), TotalSegmentator MRI (TS), and TotalVibeSegmentator (VIBE), have been proposed for multi-organ segmentation in MRI. However, the performance of these tools on specific MRI sequence types has not yet been quantified. In this work, a subset of 40 volumes from the public Duke Liver Dataset was curated. The curated dataset contained 10 volumes each from the pre-contrast fat saturated T1, arterial T1w, venous T1w, and delayed T1w phases, respectively. Ten abdominal structures were manually annotated in these volumes. Next, the performance of the three public tools was benchmarked on this curated dataset. The results indicated that MRSeg obtained a Dice score of 80.7 +- 18.6 and Hausdorff Distance (HD) error of 8.9 +- 10.4 mm. It fared the best (p < .05) across the different sequence types in contrast to TS and VIBE.
Posters - Wednesday
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View Full Details: spie.org/mi/wednesday-poster-session
Conference attendees are invited to attend the SPIE Medical Imaging poster session on Wednesday evening. Come view the posters, enjoy light refreshments, ask questions, and network with colleagues in your field. Authors of poster papers will be present to answer questions concerning their papers. Attendees are required to wear their conference registration badges.
Poster Presenters:
Poster Setup and Pre-Session Viewing: 7:30 AM - 5:30 PM Wednesday
- In order to be considered for a poster award, it is recommended to have your poster set up by 1:00 PM Wednesday. Judging may begin after this time. Posters must remain on display until the end of the Wednesday evening poster session but may be left hanging until 10:00 AM Thursday. After 10:00 AM, any posters left hanging will be discarded.
spie.org/MI/Poster-Presentation-Guidelines
Poster groupings are listed below by topic.
Posters: Abdomen
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-52
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The pre-training of transformer architectures has been shown to improve deep learning networks in various tasks including natural language processing and computer vision. While this approach has shown promise in various fields, more development and translation need to be dedicated to medical imaging applications. Current literature scarcely focuses on thorough assessment of implemented pre-training approaches as well, potentially hindering performance in downstream tasks. In this work we leverage a state-of-the-art pre-training architecture with hyperspectral imaging (HSI) to effectively encode spatial and spectral features of various ex vivo tissues. We utilize a masked autoencoding scheme to perform pre-training on an internal dataset captured with a high-speed hyperspectral laparoscopic imaging system. Pre-training results are qualitatively assessed through reconstruction visualization and quantitatively assessed with mean squared error.
13407-53
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This paper proposes a new pancreatic segmentation method, which consists of three parts. Firstly, the MFEM was proposed to address the issue of pancreas being a small organ with imbalanced categories. Secondly, the AEM is proposed to address the low contrast and blurred boundaries between pancreatic organs and surrounding tissues. It introduces deep and shallow features of reverse attention, channel attention, and asymmetric convolution fusion to further enhance the relevant target regions. Finally, we propose an edge learning branch to learn prior shape information of the organs to be segmented, in order to address the issues of large shape variations and high anatomical variability in pancreatic organs. The experimental results show that the network designed in this paper has good segmentation accuracy and stability, and has good clinical application prospects.
13407-54
The search for CT imaging subtypes of colorectal liver metastases and the impacts of slice thickness
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The potential for CT imaging subtypes of colorectal liver metastases through unsupervised analysis and the influence of slice thickness on classical radiomic, topological, and CNN-based features were explored. Preoperative, portal-venous phase, contrast-enhanced abdominal CT imaging from a multi-center cohort of 1,199 patients with resectable colorectal liver metastases were used to extract features from the liver parenchyma and largest tumor by volume. PCA and t-SNE were used to visualize clustering and slice thickness patterns. Classical radiomic first order and texture features showed specific slice thickness patterns. Topological and CNN-based features formed no visual slice thickness associations. Importantly, persistent landscape and persistent statistics features showed multiple distinct clusters, suggesting the existence of CT imaging subtypes for colorectal liver metastases. We demonstrated the importance of accounting for slice thickness when using classical radiomic features and the potential for CNN-based and topological features to be used as imaging biomarkers robust to this confounder.
13407-55
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Colorectal cancer (CRC) is a significant global health issue, responsible for numerous cancer-related deaths annually. Colorectal polyps (CRPs), as precursors to CRC, necessitate early detection and precise characterization to enhance patient outcomes and decrease mortality rates. In the realm of medical imaging, computer-assisted polyp analysis is becoming increasingly vital for screening endoscopy and diagnostic processes. To tackle the issue of class imbalance in medical image classification, this study employs a cascaded binary decision-making approach to distinguish between the different CRP pathologies. This method introduces transparency by breaking down the intricate characterization task into sequential binary decisions, reflecting a step-wise approach to decision-making and attempting to simulate a random forest-like model using neural networks. Each binary classifier focuses on discerning one class from the other classes, facilitating a more understandable decision-making process. The performance evaluations demonstrate the potential of the approach to enhance the multi-class characterization accuracy.
13407-56
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study investigates the influence of physician experience, specialty, and institutional background on the effectiveness of an AI system designed to assist in assessing treatment responses in bladder cancer. Using pre- and post-chemotherapy CTU scans from 123 patients, 17 physicians evaluated 157 lesion pairs, categorized as easy or difficult cases. Results show that AI improves diagnostic accuracy, especially in easier cases, across various physician backgrounds. The study highlights the variability in AI impact, with more significant improvements seen in experienced physicians, radiologists, and oncologists. These findings underscore the importance of tailoring the implement of AI tools to specific medical scenarios and user expertise.
13407-57
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Pancreatic ductal adenocarcinoma (PDAC) has a high rate of early liver recurrence post-surgery, significantly impacting patient survival. This study focuses on predicting early recurrence by analyzing handcrafted and deep radiomics features extracted from contrast-enhanced CT scans of 223 PDAC patients. Early recurrence was defined as cancer returning within 6 months after surgery, while non-recurrence was defined as the absence of recurrence for a minimum of 24 months. Feature selection was performed using the minimum redundancy maximum relevance (mRMR) method. Machine learning models were trained and validated, achieving AUCs of up to 0.77 for handcrafted features and 0.75 for deep features. These results demonstrate that both handcrafted and deep radiomics features provide valuable insights for predicting early recurrence in PDAC patients.
13407-58
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Detection of inflammation in MR enterography (MRE) images is essential for the diagnosis and treatment planning of inflammatory bowel diseases. However, variability in the size, location, and shape of inflammation presents challenges for automated detection systems. This often results in false positives due to the similar imaging characteristics shared between the inflammation and non-inflammation regions. In this study, we propose a novel method for detecting inflammation in MRE images by applying a context-aware Focal Modulation Network (FocalNet) to a Mask R-CNN-based approach. Unlike traditional self-attention mechanisms, the Focal Modulation Network prioritizes nearby regions and de-emphasizes distant areas. Our method integrates both visual features and distance-based contextual information, including the location of inflammation, via gating aggregation. Experimental results confirmed that the proposed method improved mAP and precision scores through false positive reduction.
13407-59
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The retroperitoneum presents diverse pathologies, including rare benign tumors and malignant neoplasms, both primary and metastatic. Diagnosing and treating these tumors are challenging due to their rarity, late presentation, and proximity to critical structures. Automatic semantic segmentation of tumors is crucial for accurate cancer diagnosis and treatment planning. Our study evaluates U-Net and its variants, including convolutional neural networks (CNNs), Vision Transformers (ViTs), Mamba, and the new xLSTM, using an in-house CT dataset. Results demonstrate the effectiveness of xLSTM within the U-Net structure, offering promising advancements in medical image segmentation. Our codes are available on GitHub.
13407-60
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Measures of kidney health are often indirect values dependent on patient demographics and need additional study. We develop deep learning models to identify patient demographic information from the kidney parenchyma. We find models that can reliably estimate patient age and gender. However, the model's ability diminishes when evaluated on a patient population with abnormal creatine levels. Our findings extend to the transplant population, where we can reliably recover information regarding the recipient with decreased performance in patients with abnormal creatine levels. Establishing patient demographic estimation could be used as an accessible surrogate for patient health.
13407-61
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Colorectal cancer is the third most common cancer worldwide. Approximately 50\% of these patients will develop liver metastases. Neoadjuvant chemotherapy is used to reduce size of metastases so that patients are eligible for potentially curative resection. Our study demonstrates that baseline CT scans can be used to predict response prior to initiation of chemotherapy. We analyzed baseline CT images of 342 patients with unresectable colorectal liver metastases (CRLM) who received chemotherapy at Memorial Sloan Kettering Cancer Center (MSK) or University of Texas MD Anderson Cancer Center (MDA). We predicted response as defined by Response Evaluation Criteria in Solid Tumors (RECIST). Different classifiers were evaluated, and the three best-performing ones were selected. Further, these were later combined in a stacking classifier to improve prediction accuracy. Naive Bayes showed the best performance with an accuracy of 0.745 and AUC of 0.785 when trained on features extracted from all tumors. The stacking classifier demonstrated a slightly better precision (0.742) and specificity (0.742).
13407-62
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Distinguishing between intrahepatic cholangiocarcinoma and hepatocellular carcinoma in imaging is a difficult task for a radiologist. We endeavoured to develop reliable models to automatically classify these tumour types. In this study we propose to use persistent homology (PH), a main tool in the field of topological data analysis (TDA) to build topological shapes from computed tomography (CT) scans of the liver. PH is used to extract topological and geometrical features from CT scans in the form of persistent barcodes. Extracted topological features are used as input to various classifiers achieving 97.56 % F1-score with 97.5 % accuracy. These results are compared with radiomics features and CNNs. Our results suggest that TDA and radiomics features can complement each other whereby a miss-classified scan by radiomics is correctly classified by TDA and vice versa.
13407-63
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal cancer with a 5-year survival rate of only 8%, largely due to late detection. This study introduces PanNet, an automatic segmentation model designed to improve PDAC segmentation over multiple datasets. PanNet introduces multi-level skip connections and a novel feature-based attention aggregation (FAA) block that enhances model accuracy by reducing false positives in tumor segmentation. The FAA block applies pixel-wise attention across all channels in 3D feature vectors in decoder blocks. The FAA block in conjunction with deep supervision significantly improves Dice Score (DSC) by up to 7.3% across two publicly available PDAC datasets. This improvement leads to a more accurate delineation of tumor margins and volumes, outperforming existing state-of-the-art pancreas segmentation models by over 60.2%. The robust performance of the proposed PanNet makes it a potential model to aid early detection of PDAC in future studies.
13407-64
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
We propose to develop deep learning models that improve the overall cancer diagnostic accuracy, by classifying radiologist-identified patients or lesions (i.e. radiologist-positives), as opposed to the existing models that are trained to discriminate over all patients. We develop a single voxel-level classification model, with a simple percentage threshold to determine positive cases, at levels of lesions, Barzell-zones and patients. Based on the presented experiments from two clinical data sets, consisting of histopathology-labelled MR images from more than 800 and 500 patients in the respective UCLA and UCL PROMIS studies, we show that the proposed strategy can improve the the diagnostic accuracy, by augmenting the radiologist reading of the MR imaging.
13407-65
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Previous study proposed Vision Retentive Network (ViR), making it a strong successor to the Vision Transformer (ViT) architecture. The ViR enabled the learning of complex spatial relationships between patch locations using chunkwise recurrent representation and a specific mask formulation based on decay strength. We developed a new ViR, named "Adaptive Region-Oriented Masked ViR" (AROMA ViR) for predicting macrovascular invasion in hepatocellular carcinoma (HCC) on CT. The model applied causal masks specialized for specific shapes of liver for each slice image into retention encoders. By guiding AI on the specific regions for training, AI enables accurate learning and provide AI decision-making rationale, thereby offering high diagnostic precision and reliability to healthcare professionals. In fact, the AROMA ViR showed better performance than the conventional model for predicting macrovascular invasion in HCC on CT image.
13407-66
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Hepatocellular carcinoma (HCC) surveillance primarily relies on ultrasound (U/S), which often exhibits decreased sensitivity in high-risk populations, such as individuals with cirrhosis or obesity. Abbreviated magnetic resonance imaging (AMRI) offers a potential alternative by employing targeted MRI sequences to enhance HCC detection. AMRI encompasses three primary strategies: non-contrast, dynamic contrast-enhanced, and hepatobiliary phase imaging, showing potential for overcoming U/S limitations in these populations. This study investigates the application of deep learning (DL) techniques to automate HCC tumor detection and segmentation within dynamic contrast-enhanced (Dyn-AMRI) protocols. Specifically, we leverage the capabilities of Vision Transformers (ViTs) to analyze complex image data and extract relevant features. Additionally, a novel heuristic is introduced to enhance the segmentation performance of the MedNeXt architecture. Our aim is to develop a robust DL pipeline for accurate HCC detection and segmentation on Dyn-AMRI, ultimately improving diagnostic outcomes.
13407-67
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Autosomal dominant polycystic kidney disease (ADPKD) is a hereditary kidney disease and is responsible for 10% of patients less than 65 years of age with end-stage kidney disease. ADPKD results in gradual enlargement of total kidney volume due to cyst growth over decades prior to decline in kidney function and kidney failure. The purpose of this work was (1) to determine whether radiomic features extracted from the non-cystic kidney parenchyma in baseline magnetic resonance imaging (MRI) scans reliably predict kidney function decline to chronic kidney disease stage 3A or greater at 60-months follow-up and (2) to determine whether radiomic features at 24-month and 48-month timepoints provide additional power in predicting kidney function decline versus baseline radiomic features alone. This is the first work to investigate the utility of MRI-based radiomic features extracted from the non-cystic kidney parenchyma in ADPKD to predict kidney function decline.
13407-68
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Colorectal cancer (CRC) has a high death rate, and its incidence rate keeps increasing. Fortunately, this cancer can be prevented through early detection and removal of adenomatous polyps. A missed polyp during a colonoscopy procedure can result in serious consequences. Hence, reducing the missed polyp detection rate is crucial to reducing CRC incidence. Polyp detectors are AI tools that can help endoscopists detect polyps. It is of utmost importance to keep improving the performance metrics of polyp detectors. Given the limited availability of polyp images for training, most detectors rely on general pre-trained models as a starting point. The choice of a pre-trained model is crucial for developing an effective polyp detector. This study explores various pre-trained models to identify the one that yields the best performance in polyp detection.
13407-69
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Pre-operative classification of ovarian tumors as benign or malignant has shown to be challenging, with current
algorithms struggling to achieve sufficient accuracy for clinical decision-making. The heterogenic nature of ovarian
tumors makes classification more difficult compared to other classification tasks such as lung nodule classification,
even for radiologists. This study compares three models across both ovarian and lung datasets, demonstrating
that while advanced deep learning approaches show promise, their performance in ovarian tumor classification
is lower compared to lung nodule classification. Our feature analysis reveals that ovarian tumors exhibit high
feature heterogeneity and lack feature robustness, underscoring the need for novel methods to enhance feature
differentiation. Our results also suggest that it is possible to characterize more difficult ovarian tumors, indicating
the potential to enhance classification accuracy through selective hard-example filtering. Further research is
necessary to break the impasse and move toward reliable clinical application in ovarian tumor classification.
13407-70
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
CT imaging is widely used to analyze pancreatic pathologies as it provides useful information regarding progression of pancreatic disease. Segmentation of the pancreas is a pre-requisite for many applications, such as longitudinal change tracking. While current approaches segment the whole pancreas and extract CT biomarkers from the segmentation, changes in pancreatic morphology and CT attenuation in certain pancreatic sub-regions may provide more details related to disease severity and prognosis. In this work, an automated 3D tool was developed to segment three main pancreatic sub-regions (the head, body, and tail) on CT, thereby allowing biomarkers to be extracted from each sub-region.
Posters: Brain
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-71
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study introduces a novel, fully automated pipeline for predicting post-stroke cognitive impairment, addressing a critical need in personalized stroke care. Our system leverages machine learning algorithms to process clinical brain scans, including anatomical and diffusion-weighted images, to identify patients at high risk of cognitive deficits. The pipeline rapidly analyzes lesion location and its overlap with critical brain regions, completing the entire process from raw DICOM data to risk prediction in under 10 minutes. Validated on a dataset of 114 acute ischemic stroke patients with comprehensive neuropsychological assessments, our results show significant differentiation between high and low-risk groups (p<0.0001). This innovation has the potential to enhance clinical decision-making, optimize rehabilitation strategies, and improve resource allocation in stroke care, paving the way for more personalized and effective treatment approaches.
13407-72
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Glioblastoma multiforme (GBM) are extremely invasive cancers. As a response to radiation treatment, in many cases, a new or a progressing lesion can be observed in imaging studies which resolves without additional treatment. This phenomenon is referred to as pseudoprogression (PsP). In contrast to PsP, a True Progression (TP) represents an enlarging lesion that requires a change in the treatment. Distinguishing between PsP and TP is thus central to treatment choice and clinical management. In this paper we present a 3D convolutional neural network (CNN) trained on 3D MRI images from 114 GBM patients to distinguish between Psp and TP. The model performs with an AUCROC of 0.74, Peak geometric mean of specificity and sensitivity: 0.69, Brier Score: 0.22, Scaled Brier Score: 0.04. Our findings suggest further investigation of deep learning models trained on larger imaging datasets to build more robust and generalizable models for distinguishing between PsP and True Progression.
13407-73
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Early prediction of Alzheimer’s Disease (AD) is crucial for optimal patient care. It can be achieved by
the inspection of suitable imaging modalities of the brain, namely structural T1-weighted MRI (T1w),
Fludeoxyglucose-18 Positron Emission Tomography (FDG-PET), and Arterial Spin Labeling (ASL). In
this work, we present image-based AD classification using a Swin Transformer model, and
investigate the effect of domain specific pretraining utilizing Masked Image Modeling. The model
was trained to predict the three classes cognitive normal (CN), mild cognitive impairment (MCI), and
AD using T1w, FDG-PET, and ASL images retrieved from the ADNI database. Our results demonstrate
the pretraining’s positive effect on the classification metrics for all modalities, reaching 92.9%. 90.3%
and 82.7% ROC-AUC. They are competitive in comparison to reported state-of-the art approaches, in
particular on the non-invasively retrieved ASL data
13407-74
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Glioblastoma is one of the most aggressive and deadliest types of brain cancer with low survival rates. Analysis of Magnetic Resonance Imaging (MRI) scans is one of the most effective methods for the diagnosis and treatment of brain cancers such as glioblastoma. Accurate tumor segmentation in MRI images is often required for treatment planning and risk assessment of treatment methods. We propose a novel pipeline, Brain Radiology Aided by Intelligent Neural NETworks (BRAINNET), which leverages MaskFormer, a vision transformer model, and generates robust tumor segmentation masks. We use an ensemble of predictions from models separately trained on each of the three orthogonal 2D slice directions (axial, sagittal, and coronal) of a 3D brain MRI volume. We train and test our models on the publicly available UPenn-GBM dataset, consisting of 3D multi-parametric MRI (mpMRI) scans from 611 subjects. Using Dice coefficient (DC) and 95% Hausdorff distance (HD) for evaluation, our models achieved state-of-the-art results in segmenting all three different tumor regions -- tumor core (DC = 0.894, HD = 2.308), whole tumor (DC = 0.891, HD = 3.552), and enhancing tumor (DC = 0.812, HD = 1.608).
13407-75
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Lesion-symptom mapping technology can uncover the relationship between the location of brain lesions and cognitive outcomes. Unfortunately, current technology can only assess a single cognitive outcome at a time, whilst patients often experience multiple cognitive problems simultaneously. This work proposes a deep learning-based approach that can predict multiple cognitive outcomes and uses eXplainable AI to generate attribution maps that pinpoint brain locations involved in the outcomes. A simulation study was used to assess the performance of two 3D deep learning approaches and two XAI techniques. From 821 patients, real WMH lesions were used as input, three ROIs were placed in the brain as impactful ground truth locations, and artificial cognitive scores were generated. Results demonstrate that the deep learning approaches can both predict the artificial scores and reconstruct the associated ROIs.
13407-76
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study aims to detect significant indicators of early AD by extracting and integrating various imaging biomarkers, including radiomics, hippocampal texture descriptors, cortical thickness measurements, and deep learning features. We analyze structural magnetic resonance imaging (MRI) scans from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohorts, utilizing comprehensive image analysis and machine learning techniques. Our results show that combining multiple biomarkers significantly improves detection accuracy. Radiomics and texture features emerged as the most effective predictors for early AD, achieving AUCs of 0.88 and 0.72 for AD and MCI detection, respectively. Although deep learning features proved to be less effective than traditional approaches, incorporating age with other biomarkers notably enhanced MCI detection performance.
13407-77
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Processing pediatric brain MR data requires specialized approaches due to the unique anatomical features of children’s brains. This study introduces a comprehensive framework for processing T1-weighted brain MR scans for children aged 2 to 8, incorporating skull stripping, registration, bias field correction, normalization, resizing, and tissue segmentation using tools like FSL, FreeSurfer, SimpleITK, and SciPy. Our framework, developed using 279 scans from unexposed controls and validated with 30 scans from children with prenatal alcohol exposure (PAE), provides tailored recommendations for each processing step. Key suggestions include FSL-BET and SynthStrip for skull stripping, age-specific templates for registration, and essential bias correction and normalization for uniform voxel intensities. While resizing and tissue segmentation are optional, FSL-FAST is recommended for non-artifact scan segmentation. The framework’s outcomes were primarily evaluated through qualitative visual inspection, supplemented by quantitative analyses in few steps, ensuring its reliability and applicability across pediatric neuroimaging studies.
13407-78
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The analysis of abnormalities in various brain regions requires combining metabolic data from PET with anatomical segmentation from MR due to the relatively low resolution of PET images. In this study, we automatically segmented two representative brain areas from hybrid PET/MR scans of twelve PD and three MSA patients using both Atlas- and DL-based methods. We then compared the Standardized Uptake Values (SUVs) and accuracy of segmentation in the corresponding regions, with manual segmentation used as the ground truth for comparison. The results of this study indicate that the DL-based method produced superior segmentation accuracy compared to the Atlas-based method. However, there were no significant differences in the SUVmax and SUVmean across the different methods for the segmentation of Caudate and Putamen. Despite being more accurate for Caudate and Putamen segmentation, the DL-based method had little effect on the calculation of their SUVs in hybrid PET/MR scans, as compared to the widely used Atlas-based method.
Posters: Breast
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-79
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Neoadjuvant chemotherapy is a systemic therapy for breast cancer. Early prediction of efficacy can help patients who will not benefit from it to make timely adjustments to their treatment regimen and reduce toxic side effects. Although numerous deep learning-based image classification methods have been developed in recent years, they often fail to effectively explore inter-modal correlations, which does not align with the clinician’s process. We propose a modal interaction attention-based multi-modal fusion network, composed of an encoder for extracting multi-sequence and cross-modal features and a decoder for fusing multi-modal features. Experiments on 214 cases of data collected from clinics demonstrate that the classification method achieves an area under the curve (AUC) of 0.898, outperforming other state-of-the-art methods. From our experimental results, the addition of modal interaction attention-based network effectively tackle the challenge of multi-modal data fusion. The proposed method has the potential to offer early prediction of NAC effects and realize more optimal treatment plans for patients.
13407-80
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Early detection of breast cancer through mammography is crucial for improving prognosis. While bilateral comparison is typically essential in mammography screening, many patients, especially post-partial mastectomy, have only unilateral imaging available. This study introduces an innovative computer-aided detection approach combining bilateral image training with single-view inference. Using a Semantic Segmentation model with ConvNextV2 backbone and an auxiliary loss to help the model learn to extract differential information between bilateral images, the Bilateral model with single-image inference (B-SI) consistently outperformed traditional single-image (S) and bilateral inference (B-BI) models in detecting mass-like lesions particularly in focal asymmetric densities (FADs) detection. This advancement enhances the ability to detect mass-like lesions in scenarios where bilateral comparison is unfeasible, potentially improving diagnostic accuracy and early detection in clinical situations.
13407-81
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this work, the effect of using radiomic information extracted from CESM studies is evaluated to quantitatively represent masses and determine its benignity or malignancy. The proposed strategy is based on the fusion of radiological texture features, the same that are extracted from amROI selected in both images (low energy ROI (LM) and recombinated ROI (CM)), respectively. Once, the characteristics are obtained, these are combined as an unique vector of relevant characteristics that integrates the information extracted from the ROIs, respectively. The relevant characteristics are selected through a main component analysis (PCA). In this work, two fusion approaches to radiological information were evaluated, in the first, the characteristics of each ROI were assembled in a single vector and then reduced, while, in the second approach, the characteristics were first reduced and then combined. Finally, the features vector obtained is used as inputs for a binary classifier, which is implemented in a vector support machine (SVM) for the classification task. The results obtained in the experiment indicated that the combination of data can exactly describe the mammographic mass featu
13407-82
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Breast cancer has the highest prevalence in the world, and thus, most countries have screening programs which
aim to detect the cancer onset early. In these screening programs, negative studies dominate the dataset. Unsu-
pervised anomaly detection promises to take advantage of the negative studies by using it to detect abnormalities
as cancer or signs of cancer onset. In this study, we evaluate an anomaly detection method for cancer predic-
tion (1-year ahead) on a MRI dataset of a high risk cohort with BRCA1 and BRCA2 gene mutations. As the
approach fails to predict cancer risk on the dataset, we investigate the intrinsic behavior of the method. Our
analysis reveals, that the reconstruction based method might only detect high intensity anomalies and that the
reconstruction quality is highly correlated with noisy patterns in the image patches.
13407-83
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Interpreting mammograms is challenging, especially for less experienced radiologists. Grading systems were introduced to categorize mammogram findings, with Category 3 and its equivalents representing particularly difficult cases due to their ambiguous nature. Artificial intelligence (AI) showed the potential to provide valuable second opinions, helping radiologists reclassify these indeterminate cases. This study analyzed data from 169 less experienced radiologists who collectively interpreted 22,200 mammograms, focusing on AI's role in improving specificity without compromising sensitivity. Various AI probability thresholds were tested to assess their impact, with lower thresholds proving most beneficial, allowing radiologists to more confidently reclassify indeterminate cases.
13407-84
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Breast cancer is the world's most prevalent cancer type. Risk models predicting the chance of near future cancer development can help to increase the efficiency of screening programs by targeting high risk patients specifically. In this study we develop machine learning models for predicting the 2 year risk for breast cancer and current breast cancer detection. Therefore, we leverage feature sets based on background parenchymal enhancement (BPE), radiomics and breast symmetry. We train and evaluate our models on longitudinal MRI data from a German high risk screening program using random forests and 5-fold cross validation. The models, which are developed similar to prior work for breast cancer risk prediction, have low predictive power on our dataset. The best performing model is based on BPE features and achieves an AUC of 0.57 for 2 year breast cancer risk prediction.
13407-85
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Distinction between invasive breast cancers and non-invasive cancers is important for determination of treatment planning. The decision is generally made based on biopsy; however, if diagnosis can be predicted by imaging, it would be useful for timely treatment planning and biopsy sampling. The purpose of this study is to classify breast ultrasound images with invasive cancers and non-invasive cancers. The number of cases used in this study is 690 breast ultrasound images, including 584 invasive cancers and 106 ductal carcinomas in situ (DCIS). Since cases are highly imbalanced, the model was first pretrained for matched pairs and unmatched pairs with contrastive loss. The model is then fine-tuned for classification of invasive cancers and DCISs. Although accuracy is almost unchanged, the recall for DCIS cases was slightly improved. Classification of non-invasive cancers on ultrasound images can support prompt treatment planning and biopsy procedures.
13407-86
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In potential individualized breast density-based breast cancer screening programs, risk-based models and when generating realistic digital breast phantoms, it is important to understand the normal changes in breast density. Breast density in 26 182 women in two consecutive screening rounds was retrospectively assessed with a commercial AI software. A small but significant decrease in breast density between two screening rounds was seen. The decline in breast density was more pronounced among women with denser breasts.
13407-87
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
We present GhostTuDNet, a compact and efficient model designed for accurate tumor detection on CPU devices. Based on GhostNetV3 and enhanced through knowledge distillation from the ViG model, GhostTuDNet achieved a Precision of 93.5% and Recall of 92.7% on the Intracranial Tumor dataset. Despite its compact size of 6.1M parameters and an inference time of 8.7ms, GhostTuDNet's performance was comparable to the much larger ViG model, which has 86.8M parameters. In contrast, MobileNetV3, another compact model, exhibited significantly lower performance. GhostTuDNet demonstrates strong potential as a solution for tumor detection in resource-limited environments, offering a balance between compactness, efficiency, and high accuracy.
13407-89
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Artificial intelligence (AI) models need to be carefully evaluated for performance on underrepresented subgroups to avoid exacerbating health disparities, but test data for such subgroups are often limited. Traditional evaluation methods often misinterpret performance differences across such limited subgroups data. We present an novel approach for meaningful subgroup analysis, based on hyperdimensional computing to encode model features during the AI model evaluation phase. The hyperdimensional representation retains the subtle subgroup characteristics and enables identification of diverging characteristics (DCs) responsible for performance differences across subgroups. Thus, we develop a technique to identify and detect these DCs and show that they reflect performance bias.
Posters: Cardiovascular
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-90
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study aimed to develop a novel machine learning model to predict PR from non-contrast CT calcium scoring (CTCS) scans. This study included 1,324 patients who underwent both CCTA and CTCS. PR was defined as an outer vessel diameter that exceeded the mean diameter of the segments immediately proximal and distal to the plaque by 10% in CCTA images. We analyzed various clinical characteristics, Agatston score (AS)-derived features, and novel epicardial adipose tissue features (fat-omics), encompassing 211 radiomic features, including morphological, spatial, and intensity parameters. We employed elastic net regression to select the most predictive features, which were then used to train a CatBoost classification model. The predictive value of our method was assessed through 1,000 repetitions of five-fold cross validation. Using the top 13 features, including 4 clinical, 3 AS-derived, and 6 fat-omics features, selected by the elastic net, our model achieved excellent classification of PR, with a sensitivity of 83.8±5.2%, a specificity of 71.1±1.2%, and accuracy of 71.9±2.0%. Among all employed methods, the CatBoost method showed the best classification results.
13407-91
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this study, we developed a novel machine learning model to predict obstructive coronary artery disease (CAD), as defined by the coronary artery disease-reporting and data system (CAD-RADS), from CTCS scans. This study analyzed 1,324 patients who underwent both CTCS and coronary CT angiography. Obstructive CAD was defined as CAD-RADS 4A-5, while CAD-RADS 0-3 were considered non-obstructive CAD. We analyzed various clinical, Agatston-score (AS)-derived, and calcium-omics features to predict obstructive CAD. The most predictive features were selected using Elastic Net regression and used to train a CatBoost machine learning model. The predictive value of the proposed method was assessed using 1,000 repeated five-fold cross-validation. Using Elastic Net, we identified the top 4 features, consisting of 1 clinical feature, 1 AS-derived feature, and 2 calcium-omics features. Our method achieved excellent classification of obstructive CAD, with sensitivity, specificity, and accuracy of 94.5±11.5%, 68.3±8.2%, and 84.1±5.9%, respectively. The inclusion of AS-derived and calcium-omics features significantly improved classification performance.
13407-92
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study explores the use of advanced AI analysis applied to low-cost (no-cost at our institution) CT calcium score exams targeting young patients not well served by conventional Agatston score (particularly Agatston<100). Using data from 16,042 patients (Agatston<100), we segmented epicardial adipose tissue and extracted 216 features and combined them with clinical features. The combined model showed significantly improved prediction of MACE over traditional Agatston score (0.74 vs. 0.55). For young patients (<52-years), an age-group specific model outperformed full-cohort models even when age was a feature, suggesting a need for specificity in modeling. This approach may provide enhanced risk assessment, particularly for younger patients, where traditional CAC scoring is less predictive.
13407-93
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Visual coronary stenosis localization and severity estimation in x-ray coronary angiography (XCA) videos is a challenging task, complicated by complex vessel structure, low image quality and heart movement. This work presents a novel workflow to automate the assessment of coronary artery disease considering multiple XCA videos with different projection angles per patient. The workflow consists of five steps for XCA video processing: selection of time frames with sufficient vessel lumen visualization, detection of stenotic regions and the corresponding coronary segment in the selected frame, calculation of the stenosis degree, movement tracking to combine detections showing the same stenosis in one video, prediction of the coronary segment and stenosis degree for the stenosis represented by a set of assigned detections. We evaluate the prediction and the corresponding degree estimation for each coronary segment on patient-level and demonstrate the impact of considering multiple projections per patient on the stenosis evaluation accuracy.
13407-94
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study presents an AI-based approach for automatically measuring heart chamber size on non-contrast CT scans using an enhanced nnU-Net segmentation model. Utilizing the OrcaScore dataset, we created paired non-contrast CT images and segmentation labels. The nnU-Net was trained and evaluated, achieving a Dice score of 0.92. Visualization of the predicted results showed excellent alignment with ground truth labels, effectively matching the heart chambers.
This advancement significantly enhances diagnostic tools and expands research opportunities in medical imaging by automating the generation of segmentation labels. Our method facilitates large-scale cardiac assessments on non-contrast CT scans, offering a method of assessing risk of cardiovascular disease.
Posters: Chest
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-95
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Guided bronchoscopy systems aid early lung cancer diagnosis, using high-resolution chest CT scans for procedure planning. Accurate disease staging requires comprehensive detection of central-chest lymph nodes, including smaller nodes (short axis ≥ 5 mm), which are often missed in manual assessments or by traditional deep learning methods using thick-slice CT scans. We propose a three-stage framework for broader size range lymph node detection on high-resolution CT scans: 1) pre-processing, 2) mode inference with a Swin-UNETR Transformer model, and 3) post-prediction filtering. Our framework showed an average sensitivity of 76.5% for detecting nodes of varying sizes across a 41-patient dataset, and outperformed state-of-the-art models, achieving a higher sensitivity of 85.4% for detecting enlarged mediastinal lymph nodes on a 30-patient public dataset.
13407-96
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Lung cancer accounts for 22% of all cancer deaths, necessitating screening with Low Dose Computed Tomography (LDCT) to improve early detection. Approximately 97% of early detected nodules are benign. Mathematical Prediction Models (MPMs) based on demographic, clinical, and radiologist interpretated data have been developed to provide insight into lung cancer risk at the time of nodule detection. However, MPMs incorporate subjective data (patient reported and interpretation). We hypothesized that a machine learning approach that utilizes only quantitative features extracted from the LDCT data, focusing on the detected nodule and surrounding lung parenchyma, can outperform existing calibrated MPMs. Our approach explores predictive performance of ensembles of Artificial Neural Networks trained on important radiomic features, showing improved specificity and AUC-pr when compared to current MPMs.
13407-97
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
According to the 2023 World Health Organization report, 7.5 million people were diagnosed with tuberculosis (TB) in 2022. TB screening often uses chest x-rays (CXRs), with significant efforts invested in automation. A key concern with algorithms that provide only image level labels is the lack of explanation for their outputs. Semantic segmentation of TB lesions can enable human supervision in the diagnosis process. This work introduces TB-Portals SIFT, a dataset for semantic segmentation of TB lesions in CXRs (6328 images with 10,435 pseudo-label lesion instances). Ten semantic segmentation models from the UNet and YOLOv8-seg architectures were evaluated in a five-fold cross validation study utilizing the new dataset. The best performing models, nnUNet (ResEnc XL), YOLOv8m-seg and their ensemble were evaluated for generalization on related classification and object detection tasks. The ensemble was found to be the most stable and best performant approach across the three tasks with a mean Dice of 0.85, AUC ranging from 0.90 to 0.98 across four sequestered datasets and a mAP@50 of 0.72 on one sequestered dataset.
13407-98
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
There is a growing awareness among radiologists of interval lung cancers (ILC), which develop after a negative screening CT scan but before the next scheduled screening. Compared to cancers detected at baseline, ILCs are often late or incurable-stage cancers with poor prognosis. We developed an AI tool that uses imperceptible texture features from whole-lung LDCT volumetric images of "normal-looking" lungs, prior to the development of an ILC, to forecast the individual risk of future cancer. We employed a convolutional neural network (CNN) algorithm to extract features describing the region of interest around developing ILCs in screening patients (Train, n=26; Test, n=10). A support vector classification (SVC) model was then trained using only the highest ranked discriminating CNN feature. The SVC model reported an archived sensitivity, specificity, and F1-score (SN: 0.82, SP: 0.82, F1: 0.75). This model has the potential to identify the development of more biologically aggressive ILCs within two years.
13407-99
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This paper addresses the problem of pathological lung segmentation, a significant challenge in medical image
analysis, particularly pronounced in cases of peripheral opacities (severe fibrosis and consolidation) because of
the textural similarity between lung tissue and surrounding areas. Fully-supervised AI models require a large
annotated training dataset to overcome the high variability between healthy and pathological lungs, which is
hardly achievable. To overcome this challenge, this paper emphasizes the use of CycleGAN for unpaired image
to-image translation, in order to provide an augmentation method able to generate fake pathological images,
matching an existing ground truth. Although previous studies have employed CycleGAN, they often neglect
the issue of shape deformation, which is crucial for accurate medical image segmentation.
13407-100
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study evaluates the performance of several pre-trained foundation models (FMs) in diagnosing tuberculosis (TB) using X-ray images. We compared three FMs trained on ImageNet, X-ray images, and 3D CT images using three publicly available TB diagnostic X-ray image collections: TBCXRay, TBX11K, and Shenzhen-Montgomery. The models extracted deep features which were then used to train a Naive Bayes classifier for TB diagnosis. The 3D CT FM achieved the highest average AUC of 0.93 on the TBCXRay validation set, while the ImageNet FM followed closely with an AUC of 0.92. However, both 3D CT and X-ray FMs exhibited higher sensitivity to batch effects, with significantly lower AUCs on the TBX11K and Shenzhen-Montgomery collections compared to the ImageNet FM. These results underscore the effectiveness of out-of-domain FMs in characterizing TB while highlighting the challenges posed by batch effects in medical imaging.
13407-102
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This work enhances lung nodule classification using the Lung Image Database Consortium (LIDC) dataset by leveraging Variational Autoencoders (VAEs) for image augmentation. Incorporating it increases sensitivity for lung nodule spiculation classification by 3.13%, demonstrating the effectiveness of VAE-based augmentation in improving computer-aided diagnosis systems.
13407-103
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In the U.S., lung cancer is the second major cause of death. Early detection of suspicious lung nodules is crucial for patient treatment planning, management, and improving outcomes. Many approaches for lung nodule segmentation and volumetric analysis have been proposed, but few have looked at longitudinal changes in total lung tumor burden. In this work, we trained two 3D models (nnUNet) with and without anatomical priors to automatically segment lung lesions and quantified total lesion burden for each patient. The 3D model without priors significantly outperformed (p < .001) the model trained with anatomy priors. For detecting clinically significant lesions > 1cm, a precision of 71.3%, sensitivity of 68.4%, and F1-score of 69.8% was achieved. For segmentation, a Dice score of 77.1 +- 20.3 and Hausdorff distance error of 11.7 +- 24.1 mm was obtained. The median lesion burden was 6.4 cc (IQR: 2.1, 18.1) and the median volume difference between manual and automated measurements was 0.02 cc (IQR: -2.8, 1.2). Agreements were also evaluated with linear regression and Bland-Altman plots. The proposed approach can produce a personalized evaluation of the total tumor burden.
13407-104
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Pneumoconiosis is an occupational lung disease caused by the inhalation of mineral dust particles. This preliminary study is among the first to explore deep learning classification classification classification of the four ordinal categories on the scale of profusion (concentration) of small opacities (0, 1, 2, or 3). We introduce hierarchical cross entropy (HCE) loss by employing a sequence of binary classification layers post-feature extraction, which ensures a more granular feature differentiation, resonating with the intrinsic ordering of severity. Comparing performance on a ResNet framework against 1) cross-entropy loss, 2) Mean-Squared Error (MSE) loss and 3) multi-task conditional loss, results show that our HCE loss obtains the highest accuracy of 71.4% on the test set, demonstrating preliminary evidence that ResNet model with hierarchical cross entropy loss can successfully be used to grade this disease.
13407-105
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study investigates improving the accuracy of rib fracture detection using modern deep learning models and highlights the thought process behind them. Despite advancements in existing models, they often lack reliability and explainability. By implementing GRAD-CAM for visualizing regions of interest and utilizing the DenseNet121 architecture, we achieved a comparable model that is both accurate and explainable. Through data preprocessing and extensive hyperparameter tuning, our model achieved 75% training accuracy and 74% testing accuracy on the CheXpert dataset with an AUROC score of 0.82 and then an 85% validation accuracy on the Medical College of Wisconsin (MCW) dataset. The study highlights the need for further improvements in model focus and localization abilities to ensure reliable automated rib fracture detection.
13407-106
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In current image-guided radiotherapy (IGRT) protocols, cone-beam CT (CBCT) scans are routinely performed on a daily or weekly basis, which is an ideal imaging modality for quantitative medical image analysis tasks, including the delta radiomics and monitoring of tumor response and anatomical change. However, compared to diagnostic fan-beam CT images, CBCT images are corrupted by significant artifacts, including streaking, cupping, shading, and scatter contamination, leading to inaccurate Hounsfield unit (HU) measurements. Such issues limit the quantitative utilization of CBCT images and impede the practice of CBCT-based radiomics. To improve the image quality of CBCT for quantitative image analysis in radiotherapy, such as segmentation and delta radiomics, we develop an unsupervised method for CBCT correction using the patient-specific score-based image prior.
13407-107
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Automated lung infected region delineation in X-ray images is essential for accurately diagnosing COVID-19 and assessing disease progression. In this paper, we introduced the COVID-19 Infection Segmentation Vision Language Model (COVSeg-VLM) network for infected region segmentation in 2D X-ray images. Although the advancements in vision deep-learning neural networks have significantly enhanced the auto-contouring of normal organs and anatomical structures, accurate delineation of infected region of COVID-19 remains a challenge. Our model adeptly incorporates large language models to extract text-rich features from clinical reports, captures both vision- and text-contextual information for accurate segmentation. We assessed our model using the QaTa-COV19 Dataset, showcasing its superior performance compared to conventional models, especially in difficult cases. The results underscore our network's robustness and precision, validating its efficiency in both typical and complex segmentation scenarios. These improvements herald a new era of reliable and accurate medical image analysis, with promising implications for enhancing clinical outcomes.
Posters: Classification
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-108
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Deep Learning (DL) comprehends methods to enhance medical image classification and help physicians speed
up diagnosis. However, these methods present security issues and are vulnerable to adversarial attacks that result in the model’s misclassification, presenting severe consequences in the medical field. We propose SentinelAdvMedical, a novel pipeline to detect adversarial attacks by employing controlled Out-of-Distributions (OOD) strategies to enhance the “immunity” of DL models. Towards that end, we studied the classification of Optical Coherence Tomography (OCT) images of Skin lesions with ResNet50. Our findings show that the best OOD detectors for OCT and Skin Lesion datasets are MaxLogits and Entropy, which outperform baselines Maximum Softmax Probabilities (MSP) and Mahalanobis feature-based score. To conduct this study, we developed a novel pipeline and studied the application of OOD strategies against adversarial examples, aiming to detect them and provide security specialists with a path to check possible attacked spots in medical datasets employing the best OOD detectors in these settings.
13407-109
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Myelodysplastic syndrome (MDS) is a clonal pathology affecting hematopoietic stem cells, leading to dysplasia and cytopenia on the complete blood count (CBC). MDS suspicion is established using the patient’s complete blood count (CBC) to count the number of different cell types, followed by a blood smear examination to detect any abnormal cells. While CBC analysis is widely automated nowadays, the blood smear examination process is still mainly realized manually by expert cytologists. In this work, we propose a deep learning framework leveraging multi-task learning and multi-view convolutional neural networks in order to extract deep representation of global cell dysplasia in patients. We then combine this deep latent space information of the network with CBC parameters to perform multimodal prediction of the MDS among patients. To train our framework, we gathered a multicentric dataset of patients labelled as MDS (n=60) or control (n=57), with both blood cells images and CBC parameters per sample. Our multimodal framework outperforms CBC-based, state-of-the-art methods for MDS diagnosis on this data with an accuracy of 79.55%.
13407-110
Obesity prediction from structural MRI using conformal deep learning with uncertainty quantification
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Traditional obesity prediction studies often rely on magnetic resonance imaging (MRI) voxel-based morphometry to correlate BMI with obesity-related clinical measurements and brain structure, predominantly grey matter volume (GMV). However, these studies often show inconsistencies in regional GMV findings among obese patients, making it challenging to establish a clear relationship between obesity and brain structure. To address these limitations, we propose a computational model to predict obesity directly from individual T1-weighted structural MRI data. The proposed conformal deep learning model achieves a 5-fold cross validated average precision of 77.65% and an F1-score of 75.42%, effectively predicting the probabilistic outcome of obesity from structural MRI. Furthermore, our model provides probabilistic uncertainty quantification paired with gradient-based localization maps that discover key brain regions, such as lobes and white matter tracts for obesity prediction.
13407-111
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Medical image analysis has emerged as critical research domain because of its usefulness in different clinical applications, such as early disease diagnosis and treatment. Convolutional neural networks (CNNs) have become standard in medical image analysis due to their superior ability to interpret complex features, often outperforming humans. In addition to CNNs, transformer architectures also have gained popularity for medical image analysis tasks. However, despite progress in the field, there are still potential areas for improvement. This study evaluates and compares both CNNs and transformer-based methods, employing diverse data augmentation techniques, on three medical image datasets. For Chest X-ray, the vision transformer model achieved the highest F1 score of 0.9532 and Matthews correlation coefficient (MCC) of 0.9259. Similarly, for the Kvasir dataset, we achieved an F1 score of 0.9436 and MCC of 0.9360. For the Kvasir-Capsule, the ViT model achieved an F1-score of 0.7156 and an MCC of 0.3705. We found that the transformer-based models were better or more effective than various CNN models for classifying different anatomical structures, findings, and abnormalities in medic
13407-112
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Topological data analysis (TDA) uncovers crucial properties of objects in medical imaging. Methods based on persistent homology have demonstrated their advantages in capturing topological features that traditional deep learning methods cannot detect in both radiology and pathology. However, previous research primarily focused on 2D image analysis, neglecting the comprehensive 3D context. In this paper, we propose an innovative 3D TDA approach that incorporates the concept of superpixels to transform 3D medical image features into point cloud data. By utilizing Gaussian Optimized Coefficient, the proposed 3D TDA method, for the first time, efficiently generate holistic Persistence Images for 3D volumetric data. Our 3D TDA method exhibits superior performance on the MedMNist3D dataset when compared to other traditional methods, showcasing its potential effectiveness in modeling 3D persistent homology-based topological analysis when it comes to classification tasks.
13407-113
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Accurate determination of burn wound depth is crucial in the resection of injured tissue to avoid infection, complications in healing, and the unnecessary removal of healthy tissue. This study aims to lay the groundwork for improving burn depth classification accuracy through multispectral short-wave infrared (SWIR) imaging and deep learning. 267 regions of interest (ROIs) were imaged at wavelength bands ranging from 1200 to 2250 nm. Burn categories were blindly classified by surgeons using visual light images of the ROIs. When trained on the SWIR images alone, a MobileNet CNN model showed average accuracies ranging from 0.18 to 0.94 in predicting the surgeons’ classification of operable burns, superficial thickness burns, and normal skin. Classification accuracy for operable burns is significantly higher (up to 94%) compared to other classes. Ongoing work to augment accuracy includes incorporating spectral and texture-based features, histological classification, and segmentation of burns.
Posters: Head, Neck and Eye
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-114
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
We introduce a novel Generative Adversarial Network (GAN)-based normative modeling technique for analyzing MRI derived brain measures and detecting heterogeneous effects of brain disorders. The proposed method learns to synthesize a patient-specific control by removing disease-related variations from brain measures while preserving disease-unrelated variations. The deviation of the patient from the synthesized control, acts as a personalized image-derived biomarker that is sensitive to disease effects and their severity. In the current study, we demonstrate the method's utility by applying it to detect deviations in brain measures derived from two different imaging modalities: 1) using structural MRI derived brain measures to detect neuroanatomical deviations in participants with Alzheimer's disease, and 2) using functional MRI derived brain measures to detect functional connectivity deviations in participants exposed to traumatic brain injury.
13407-115
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Accurate early diagnosis of ocular diseases like glaucoma and age-related macular degeneration (AMD) is essential to prevent irreversible vision loss. Our novel Cross-Attentive Prototypical Few-Shot Learning (CAP-FSL) model addresses the challenges of deep learning in Optical Coherence Tomography (OCT) classification, including labeled data scarcity and the complexity of 3D analysis. CAP-FSL uses a 3D CNN backbone with a cross-attention mechanism to refine feature embeddings dynamically, enhancing inter-class discrimination and intra-class compactness. Evaluated on Duke and POAG datasets, CAP-FSL outperforms other few-shot learning methods in accuracy, sensitivity, and specificity for glaucoma, AMD, and normal tissue detection, offering a transformative solution for early ocular disease detection.
13407-116
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Retinopathy of prematurity (ROP) is a disease of the developing retina that can potentially lead to blindness in prematurely born infants. Detection of ROP in its early stages is crucial for timely treatment and has motivated the development of deep learning models to detect ROP and/or to stage its severity from fundus images. However, as in most applications of pediatric medical image analysis, the availability of labeled training data for this task is usually highly limited. We, therefore, test whether fine-tuning a publicly available foundation model for adult retinal images (RETFound) for ROP staging results in accuracy benefits over using generic image classification models. We perform extensive experiments on the largest publicly available ROP dataset and surprisingly find that RETFound, despite having seen nearly one million adult fundus images during pre-training, does not outperform the generic models pre-trained on fully task-unspecific natural images from ImageNet.
13407-117
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Otoscopy, visual examination of the ear, is essential for diagnosing middle ear infections, but accurate diagnosis requires a degree of clinical expertise. Automated deep learning methods have been explored to aid in diagnosis, typically relying on selected otoscope images. However, in real-world scenarios, clinical expert selection is not always feasible, and variations in brightness and color pose challenges. We propose a video classification approach using VideoMAE, leveraging all frames in a video to improve accuracy and eliminate the need for human selection. Our study developed a deep learning model to classify otoscope videos for seven middle ear conditions. Data augmentation techniques enhanced model generalization, and we resampled videos for balanced representation. The model, trained with 224x224 pixel videos with 16 frames each, optimized parameters to minimize classification errors. Results showed improved accuracy with data augmentation and balanced resampling, achieving 92.1%±4.0% accuracy.
13407-118
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Quality and completeness of images taken during endoscopy of the esophagus are of high importance for their use in computer-aided detection (CADe) systems. Providing a tissue map to show visited locations with sufficient image quality, and providing feedback on the completeness of the procedure could ensure standards are met for input to CADe systems. However, existing solutions are either not translatable to esophageal endoscopy procedures or impose unrealistic restrictions on endoscope movement. Our study shows the first steps in developing a tissue mapping algorithm for esophageal endoscopy, that is robust against camera rotation and movement natural to endoscopy procedures. The method combines ORB-feature-based visual odometry and image stitching to generate a 2.5D map of the esophagus. The algorithm is tested in a virtual model environment, on 4 scenarios mimicking the movements and rotations of an endoscope during a pullback procedure. Results are evaluated based on ground-truth camera pose and distortion of the generated map. It is demonstrated that the proposed algorithm can generate accurate tissue maps but is sensitive to errors in estimated camera position and rotation.
13407-119
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Vision transformers generate both high accuracy and pathology related attention maps for clinical decision. We evaluated the performance of DeiT for glaucoma diagnosis with different training methods and inputs. Retia layers thickness maps and enface images were compared for pathology pattern, while transfer learning vs. re-training was tested for attention maps patterns. For example, the pathological change of glaucoma is regional RNFL thinning and thus the retraining with thickness maps yielded superior diagnostic accuracy. Other vascular disease might benefit from transfer learning with enface images. DeiT outperformed other benchmark models with the distilled superior small objects detection.
13407-120
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study presents a novel approach for predicting survival outcomes in head and neck cancer (HNC) patients by integrating Computed Tomography (CT) and Positron Emission Tomography (PET) imaging data within a Vision-Graph Neural Network (Vision-GNN) architecture. By combining the anatomical details from CT with the functional insights from PET, the model provides a comprehensive analysis of the tumor environment, significantly enhancing the accuracy of survival predictions. The model was trained and validated using the "Head-Neck-PET-CT" dataset from the Cancer Imaging Archive (TCIA), achieving an accuracy of 0.8590, and a ROC-AUC score of 0.9088. These results demonstrate the efficacy of our approach in enhancing survival predictions for HNC patients.
13407-121
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Periodontal inflammatory diseases pose a global health challenge as they lead to tooth loss and are linked with systemic diseases. Affecting 7.4% of the world’s population, severe periodontitis requires early diagnosis and lifelong monitoring. Traditional diagnostic methods are time-consuming and subjective. Hyperspectral imaging (HSI) offers a new approach by capturing spatial and spectral data to assess inflammation. This study, conducted at TU Dresden, involved 93 participants and used HSI and AI algorithms to classify periodontal disease. The AI achieved 94.6% specificity and 84% sensitivity, marking a promising first step toward clinical application, though further research and development are needed.
Posters: Methods
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-122
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Our paper introduces GS-TransUNet, an innovative approach that combines 2D Gaussian splatting with Transformer UNet architecture to concurrently optimize skin lesion segmentation and classification. Unlike traditional models that rely on per-pixel segmentation masks, GS-TransUNet employs 2D Gaussian splatting for mask rendering, enhancing segmentation consistency and classification accuracy.
Evaluated on the ISIC-2017 and PH2 datasets, GS-TransUNet demonstrates superior performance over current state-of-the-art models in key metrics. This success in leveraging 2D Gaussian splatting and exploiting skin lesion structure marks a significant advancement in automated skin cancer diagnosis. Our approach encourages further research into integrated medical image analysis methods, potentially reshaping computer-aided diagnostic techniques in dermatology.
13407-123
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Artificial Intelligence (AI) tools have become increasingly prevalent in all aspects of healthcare. However, their adoption into the clinical setting has been limited. Monitoring a clinically deployed device to detect performance drift is an essential step to ensure the patient safety and effectiveness of the device. In this work, we describe a statistical tool for AI monitoring using cumulative sum (AIM-CU) control chart. AIM-CU computes: (i) the parameter choices for change-point detection based on an acceptable false alarm rate (ii) detection delay estimates for a given displacement of the performance metric from the target for those parameter choices.
13407-124
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Federated learning is an important framework in medical imaging AI due to the ability to access additional data in model development without the need for communication of protected information. In this study, we evaluated the robustness of data characteristics, including demographic, imaging, and class variables, in unbalanced federated learning scenarios. We trained ResNet50 models in 3 different data distribution scenarios for the task of COVID-19 diagnosis on chest X-ray images: 1) centralized learning, 2) federated learning with balanced data, and 3) federated learning with varying levels of unbalanced data. Preliminarily, we found that imaging and class variables were more sensitive to performance variation across unbalanced federation sites, and identified which variables should be prioritized by model developers.
13407-125
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Effective Parameter-Efficient Fine-Tuning (PEFT) of segmentation models for computed tomography (CT) scans is crucial for liver tumour treatment planning, including preoperative volumetrics and radioembolization dose calculations. This study integrates Low-Rank Adaptation (LoRA) into the Swin UNETR, the state-of-the-art (SOTA) abdominal segmentation model, to enhance fine-tuning efficiency with minimal CT data for liver and tumour segmentation. Evaluating various LoRA ranks and sample sizes, we found that a rank of 8 achieved optimal performance, improving the Dice score by 15% and reducing the 95th percentile Hausdorff distance (HD95) by 66% compared to traditional fine-tuning, which fine-tunes the full model's parameters. Our LoRA-enhanced model demonstrates high adaptability with limited data through a few-shot learning (FSL) approach, providing a robust solution for efficient fine-tuning of deep learning (DL) segmentation models in clinical settings.
13407-126
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
As a representation of machine intelligence, current image-driven machine learning (ML) has made significant strides in emulating the expertise of medical professionals in interpreting medical image characteristics to predict lesion malignancy. Since experts’ interpretations of medical images do not constitute the tissue pathology or the ground truth of malignancy or benignity for lesion diagnosis, these ML algorithms perform comparably with the experts but do not exceed significantly their performance. We hypothesize that this limitation stems from the ML algorithms' lack of integration of prior knowledge beyond the data they analyze. This study aims to test this hypothesis by leveraging prior knowledge in machine intelligence (pkMI) to improve significantly the performance of specific tasks for which the data are acquired, such as low dose computed tomography (CT) screening for early lung cancer.
13407-127
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Deep learning enhances medical imaging by improving diagnostic accuracy and speed, but privacy regulations often limit data availability for training robust models. Federated learning (FL) addresses this by enabling collaborative model training across institutions without compromising privacy. This study compares FL and centralized learning (CL) in medical image classification and segmentation tasks. Using chest X-ray images for classification and cardiac CT images for segmentation, experiments evaluated FL with varying client setups. FL models consistently showed high classification accuracy and high Dice similarity coefficient for segmentation, closely matching CL performance. These results indicate FL as a competitive, privacy-preserving alternative to CL in medical imaging.
13407-128
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Traditional bias correction methods in medical imaging, whether through post-training adjustments, data pre-processing, or in-training techniques, typically rely on labeled information such as age, gender, or ethnicity. This reliance may lead to overlooking subtle biases that arise from less obvious data variations or demographic factors and fails to utilize clinically relevant information that may not be labeled. Our proposed approach addresses these gaps by integrating optimal transport (OT)-based bias mitigation directly into the AI model training process without the requirement of labeled information. This method measures and aligns feature distributions within a class by leveraging their inherent similarities, capturing nuanced discrepancies that conventional approaches might miss.
13407-129
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study investigates how different tints affect the extraction of radiomic texture features from ultrasound (US) images. Using data from a Philips Imaging unit and a CIRS phantom, four tints were applied to regions of interest with varying echogenicity. The analysis revealed that tints can influence the extraction of texture features, with certain tints affecting texture patterns more significantly, particularly at low contrast levels. Overall, the study highlights the potential significance of understanding the impact of tint on computer-aided diagnosis tasks and methods.
13407-130
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Studies have noted limited generalizability of deep learning models for medical image segmentation under domain-shift. Do the models generalize well under controlled domain-shift resulting from different acquisition protocols or varying modalities? This is an important question, particularly, while segmenting tumors from a variety of cancer types to provide automation in a subsequent tumor analysis pipeline for applications such as tumor burden assessment and quantitative biomarker extraction. For tumor analysis, positron emission tomography/computed tomography (PET/CT) is a popular modality used across different cancer types. In a publicly available dataset of PET/CT images acquired using a well-defined protocol on a single scanner (controlled domain-shift), we explored the variability of segmentation across lung cancer, melanoma, and lymphoma, using two state-of-the-art deep learning models, namely nnU-Net and Segment Anything Model (SAM). Our results highlight significant variation across cancer types and noted more reliable performance in lung cancer. Our study emphasizes on taking caution while incorporating automated tumor segmentation in tumor analysis pipelines.
13407-131
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This paper proposes an interactive 3D medical image segmentation method that can deal with limited training dataset. The object boundary prompt simulation strategy and the loss function are carefully designed to train the model. The experimental results show that the performance of the interactive segmentation method remarkably surpasses that of the fully automatic segmentation method, which demonstrates the significant effects of prompt engineering. The experiments also illustrate the moderate costs of applying the interactive segmentation method. In summary, interactive segmentation methods can significantly reduce the time and labor consumed for acquiring precise annotations on the condition of lacking training materials, achieving a better trade-off between model performance and human effort for training the model.
Posters: Musculoskeletal
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13407-132
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Osteopenia, a condition marked by reduced bone mineral density (BMD), affects millions globally. Patients diagnosed with osteopenia typically have a BMD valued below the normal reference measure, but not low enough to meet the diagnostic criteria for osteoporosis. In this work, we fine-tuned deep network models on a pediatric wrist X-ray dataset to classify osteopenia versus healthy bone patterns. We also employed a decision interpretability algorithm to understand the CNNs' predictions. Our results indicate that CNNs can effectively learn bone tissue patterns, achieving an accuracy of over 95% in classifying osteopenia in pediatric wrist X-rays. The visual explanations may offer valuable insights into the key areas within the X-ray images that contribute most to the deep network predictions.
13407-133
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Multi-Task Transfer Learning Based on Multi-View MRI Images for Diagnosis of Patellar Instability
Authors
Qin Zhang,Yao Lu
Abstract
Patellofemoral Instability (PI) is a condition where the patella cannot slide normally within the femoral trochlea, leading to a range of symptoms. Diagnosis primarily relies on medical imaging conducted by physicians; however, imaging-based diagnosis still requires significant manual intervention, making the diagnostic process cumbersome and complex. Recently, there has been no computer-aided system directly diagnosing PI, making it both important and urgent to develop such systems to improve diagnostic efficiency. MRI images, with their excellent tissue contrast, provide comprehensive information necessary for assessing bone and muscle lesions, which is crucial for diagnosing PI. Therefore, we established a deep learning model based on multi-view MRI images. However, due to the relatively scarce dataset for PI, the model’s performance is suboptimal. To address the data limitations and enhance model performance, we propose a multi-task transfer learning approach.
13407-134
19 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Bankart lesions, or anterior-inferior glenoid labral tears, are diagnostically challenging due to their subtle imaging features and overlap with normal anatomy. Standard MRIs often fail to provide sufficient contrast for reliable detection, leading to reliance on invasive MRI arthrograms (MRAs). This study develops deep learning (DL) models to detect Bankart lesions on both standard MRIs and MRAs, aiming to reduce dependence on MRAs.
We curated a dataset of 586 MRIs (335 standard, 251 MRAs) with intraoperative findings (diagnostic gold standard) as the ground truth. Separate models were trained for each modality using the Swin Transformer architecture, with predictions from sagittal, axial, and coronal views combined via ensembling to maximize performance.
Our models achieved an AUC of 0.87 (83% sensitivity, 86% specificity) on standard MRIs and 0.90 (82% sensitivity, 86% specificity) on MRAs. These findings demonstrate DL’s potential to improve diagnostic confidence and reduce the need for invasive imaging.
Thursday Morning Keynotes
20 February 2025 • 8:30 AM - 10:00 AM PST | Town & Country B/C
Session Chairs:
Susan M. Astley, The Univ. of Manchester (United Kingdom), Andrzej Krol, SUNY Upstate Medical Univ. (United States)
View Full Details: spie.org/thursday-morning-keynotes
8:30 AM - 8:35 AM:
Welcome and introduction
8:35 AM - 8:40 AM:
Award announcements
- Robert F. Wagner Award finalists for conferences 13407 and 13410
- Computer-Aided Diagnosis Best Paper Award
13407-508
Pioneering vision: The journey of building AI for transformative healthcare
(Keynote Presentation)
20 February 2025 • 8:40 AM - 9:20 AM PST | Town & Country B/C
13410-509
Machine learning in neuroimaging: Understanding heterogeneity of neurologic diseases and building personalized imaging-AI biomarkers
(Keynote Presentation)
20 February 2025 • 9:20 AM - 10:00 AM PST | Town & Country B/C
Show Abstract +
Machine learning has transformed medical imaging in general, and neuroimaging in particular, during the past two decades. We review our work in this field, starting with early contributions on developing personalized predictive markers of brain change in aging and Alzheimer’s Disease, and moving to recent weakly-supervised deep learning methods, aiming to dissect heterogeneity of brain change in neurodegenerative and neuropsychiatric diseases, as well as in brain cancer. We show that disease-related brain changes can follow multiple trajectories and patterns, which have distinct clinical and genetic correlates, thereby suggesting a dimensional approach to capturing brain phenotypes, using machine learning methods.
Break
Coffee Break 10:00 AM - 10:30 AM
Session 10:
Brain
20 February 2025 • 10:30 AM - 12:30 PM PST | Town & Country C
Session Chairs:
Chuan Zhou, Michigan Medicine (United States), Hiroyuki Yoshida, Massachusetts General Hospital (United States), Harvard Medical School (United States)
13407-42
20 February 2025 • 10:30 AM - 10:50 AM PST | Town & Country C
Show Abstract +
In the U.S., the Food and Drug Administration (FDA) has recently approved over 100 devices with AI capability. Research breakthroughs in AI have led to a corresponding sharp rise in patenting activity worldwide. In the future, foundation models will provide an initial starting point to finetune models for different downstream tasks. Even so, fine-tuning foundation models is challenging due to their large number of parameters, limited availability of neuroimaging data sets for fine-tuning, coupled with limited compute resources. In this work we test different parameter-efficient finetuning (PEFT) methods to greatly reduce the total number of trainable parameters for multiple neuroimaging tasks. We show that PEFT methods can be competitive with and outperform full fine-tuning in test performance with a significant reduction in model parameters (0.04 to 32%) across multiple tasks. PEFT methods boosted performance in resource constrained settings, using only 258 MRI scans, by 3% for AD classification.
13407-43
20 February 2025 • 10:50 AM - 11:10 AM PST | Town & Country C
Show Abstract +
Alzheimer's Disease (AD) is a progressive neurodegenerative disorder impacting global health. Amyloid beta protein (Aβ) is an important biomarker for Alzheimer's disease, with levels rising early in the disease course. Patients can be categorized into normal and abnormal regarding amyloid status. Since measuring Aβ requires invasive procedures, we present a method to predict amyloid state using multi-modal medical image data with transformer-based machine learning architectures. Data from ADNI included structural T1-weighted MRI and FDG-PET scans. We compared two architectures for amyloid status classification and AD vs. MCI (mild cognitive impairment) vs. CN (cognitively normal) classification. In our best setup, we achieved 91.34% (±3.62%) accuracy for predicting amyloid state and 91.83% (±1.10%) accuracy for distinguishing between AD, MCI, and CN. Our method outperformed the best mono-modal setup (73.67% ±4.70% on T1-weighted MRI) and showed competitive performance against a literature reference using FDG-PET. Our results highlight the potential of using non-specific amyloid-related data, like T1-weighted MRI and FDG-PET, for accurate amyloid state prediction.
13407-44
20 February 2025 • 11:10 AM - 11:30 AM PST | Town & Country C
Show Abstract +
we propose a personalized approach with image-guided computational model (digital twin) that incorporate physics-based modeling to predict tumor recurrence. Our digital twin involves an inverse modeling step, followed by a recurrence model that accounts for varying surgical effects. The physics-guided inverse model considers discrete loss, and estimates patient-specific diffusion (D) and proliferation (ρ) parameters from pre-operative magnetic resonance imaging (MRI) of 133 patients from publicly available TCIA dataset. The proposed model is personalized due to use of the patient-specific parameters gleaned from the real patient data to assess risk for both high-aggressive and low-aggressive tumor groups. The prognostic index for each patient reveals the interplay between tumor aggressiveness, surgical resection, and survival outcome.
13407-45
20 February 2025 • 11:30 AM - 11:50 AM PST | Town & Country C
Show Abstract +
Medulloblastoma is the most common malignant brain tumor of central nervous system in children. Despite advances in treatment, survivors often face severe long-term side effects. Definitive diagnosis of medulloblastoma currently requires a surgical specimen for integrated histopathological and molecular analysis. This study enhances the diagnosis and molecular classification of medulloblastoma using non-invasive machine learning on MRI-derived features. The model achieved over 85% balanced accuracy in medulloblastoma differentiation and over 80% in subtype classification. Features were ranked by their importance for both differential diagnosis and molecular subtype classification, showcasing machine learning's potential for improving diagnostic precision and personalizing treatment.
13407-46
20 February 2025 • 11:50 AM - 12:10 PM PST | Town & Country C
Show Abstract +
We introduce large-scale Non-Linear Granger Causality (lsNGC) to analyze effective connectivity in high-dimensional systems like fMRI data. By incorporating dimension reduction into a non-linear prediction approach, lsNGC enables directed, multivariate causality analysis in complex networks. Applied to resting-state fMRI data from the COBRE repository, lsNGC classifies schizophrenia effectively. Utilizing brain connections as features, we used support vector machines for classification, with lsNGC significantly outperforming existing methods. Achieving a mean AUC of 0.876 and an F1-score of 0.831, it surpasses other techniques like correlation, local models, and antisymmetric correlation in statistical tests. This highlights lsNGC's potential as an imaging biomarker for schizophrenia and its clinical applicability in fMRI-based disease classification.
13407-47
20 February 2025 • 12:10 PM - 12:30 PM PST | Town & Country C
Show Abstract +
We propose a diffusion model based on Worley-Perlin noise, which provides higher robustness for image synthesis problems with striated textures, e.g. subarachnoid hemorrhage (SAH) CT scans. Unlike Gaussian noise, which uniformly perturbs images and poorly affects low-frequency regions, or Simplex noise, which excessively disrupts images, leading to distortion, our method considers the morphological structure of textured images. It ensures effective perturbation while avoiding distortion of generated images, and generates high-quality samples to tackle downstream work. To validate the effectiveness of our method, we conducted experiments on SAH CT detection with imbalanced data. The experimental results demonstrated that our synthesized data can help improve the classification accuracy in an imbalanced data situation, with the F1-score improving by nearly 6.4\%.
Break
Lunch Break 12:30 PM - 1:50 PM
Session 11:
Abdomen II
20 February 2025 • 1:50 PM - 3:10 PM PST | Town & Country C
Session Chairs:
Xiaofeng Yang, The Winship Cancer Institute of Emory Univ. (United States), Zhengrong (Jerome) Liang, Stony Brook Univ. (United States)
13407-48
20 February 2025 • 1:50 PM - 2:10 PM PST | Town & Country C
Show Abstract +
Pheochromocytoma and Paraganglioma (PPGL) tumors are uncommon neuroendocrine tumors that exhibit substantial genetic and phenotypic variability, making their classification challenging. We describe a new method
that performs supervised contrastive learning leveraging DinoV2 feature representations to categorize genetic
clusters of PPGL tumors into four distinct subtypes: SDHx, VHL-EPAS1, kinase signaling, and sporadic. The
proposal was tested on a dataset consisting of 650 contrast-enhanced computed tomography (CE-CT) axial images
of 287 PPGL patients. The results showed a balanced accuracy of 0.582±0.079, an AUC score of 0.805±0.055,
and an F1 score of 0.551±0.063 using five-fold cross-validation. Our methodology demonstrated superior performance compared to other methods by balancing the performance across all four genetic types, achieving a
2% higher balanced accuracy and a 7–14% higher F1 score, while using the entire axial slice as a single sample
in contrast to using the extracted tumor patches.
13407-49
20 February 2025 • 2:10 PM - 2:30 PM PST | Town & Country C
Show Abstract +
Pancreatic cancer is a uniformly deadly disease. Prediction of tumour response to neoadjuvant therapy is critical in determining which patients should undergo invasive surgery. Non-invasive biomarkers of response would address gaps in the management of patients. This study employs pre-trained convolutional neural networks (CNNs) to predict response from baseline computed tomography (CT) scans alone, prior to neoadjuvant therapy. ResNet50, InceptionV3, VGG16, and Xception were trained on a dataset of annotated CT scans of patients with pancreatic ductal adenocarcinoma (PDAC). Our results demonstrate that the ResNet50 model achieves the highest performance among the models predicting response, with an average (average ± standard error of the mean at 95% confidence level) accuracy of 0.679 ± 0.057, an F1-score of 0.665 ± 0.072, recall of 0.717 ± 0.081, precision of 0.698 ± 0.074, and an area under the receiver operating characteristic curve (AUC-ROC) of 0.781 ± 0.162 across 5-fold cross-validation. These findings highlight the potential for non-invasive imaging biomarkers in predicting response to neoadjuvant therapy in PDAC.
13407-50
20 February 2025 • 2:30 PM - 2:50 PM PST | Town & Country C
Show Abstract +
We propose Dynamically Calibrated Convolution (DCC), Dynamically Calibrated Downsampling (DCD), and Dynamically Calibrated Upsampling (DCU) modules, respectively. The DCC module can utilize global inter-dependencies between spatial and channel features to calibrate these features adaptively. The DCD module enables networks to adaptively preserve deformable or discriminative features during down- sampling. The DCU module can dynamically align and calibrate up-sampled features to eliminate misalignments before concatenations. We integrated the proposed modules into a standard U-Net, resulting in a new architecture, termed Dynamic U-Net. This architectural design enables U-Net to dynamically adjust features for different organs. We evaluated Dynamic U-Net in two abdominal multi-organ segmentation benchmarks. Dynamic U-Net achieved statistically improved segmentation accuracy compared with standard U-Net.
13407-51
20 February 2025 • 2:50 PM - 3:10 PM PST | Town & Country C
Show Abstract +
An accurate segmentation of the pancreas on CT is crucial to identify pancreatic pathologies and extract imaging-based biomarkers. However, prior research on pancreas segmentation has mainly focused on modifying the segmentation model architecture or utilizing pre- and post-processing techniques. In this article, we investigate the utility of anatomical priors to enhance the segmentation performance of the pancreas. Two 3D full-resolution nnU-Net models were trained, one with 8 refined labels from the public PANORAMA dataset, and another that combined them with labels derived from the public TotalSegmentator (TS) tool. The addition of anatomical priors resulted in a 6% increase in Dice score (p < .001) and a 36.5 mm decrease in Hausdorff distance for pancreas segmentation (p < .001). Moreover, the pancreas was always detected when anatomy priors were used, whereas there were 8 instances of failed detections without their use. The use of anatomy priors shows promise for pancreas segmentation and subsequent derivation of imaging biomarkers.
Program Committee
UCLA Ctr. for Computer Vision & Imaging Biomarkers (United States)
Program Committee
Peter L. Reichertz Institut für Medizinische Informatik (Germany)
Program Committee
Washington Univ. School of Medicine in St. Louis (United States)
Program Committee
Fraunhofer-Institut für Digitale Medizin MEVIS (Germany), Jacobs Univ. Bremen (Germany)
Program Committee
Massachusetts General Hospital (United States), Harvard Medical School (United States)
Program Committee
Massachusetts General Hospital (United States), Harvard Medical School (United States)
Related Courses
SPIE Courses at Medical Imaging
-
See courses on topics related to Computer Aided Diagnosis.
-
Pre-registration is required for SPIE Courses.