Deep learning classification of en face multi-spectral optical coherence tomography images

Colorectal cancer (CRC) is one of the top causes of malignancy. Although screening has significantly reduced CRC mortality, colonoscopy suffers from inadequate sampling of the tissue but, also, high cost, since many polyps that are resected and processed are benign. To alleviate this burden on the healthcare system, a “leave-in-situ” strategy is being considered. To employ such a strategy, highly accurate, in vivo, evaluation of the polyps is required, a need which can be addressed by Optical Coherence Tomography (OCT). However, thus far, OCT studies focusing on morphological features have not been proven effective. To improve the classification, the spectral information of the OCT interferogram can be exploited. In this study, we propose the use of multi-spectral analysis of en face OCT images, i.e., the utilization of images created from different bands of the spectrum, to classify human colon polyps as benign or malignant potential. Multiple, narrow-band, images at different center wavelengths and their ratios were created from 143 samples and ranked according to their predictive potential. The top three spectral features were combined to create a color multi-spectral image. The multi-spectral images were subsequently classified using a DenseNet deep learning model. The resulting accuracy, sensitivity and specificity were 91%, 94%, and 88% respectively. The proposed approach must be expanded to include more polyps and explore more sophisticated multi-spectral methods. However, these preliminary results provide evidence that this method has the potential to improve the accuracy of OCT and, in the future, enable “leave-in-situ” approaches.