Multitask transfer learning based on multiview MRI images for diagnosis of patellar instability

Patellofemoral instability (PI) is a condition characterized by the inability of the patella to glide normally within the femoral trochlear groove, leading to a range of symptoms. Diagnosis primarily relies on medical imaging interpreted by physicians, which involves substantial manual intervention and complex workflows. Currently, no computer-aided systems specifically target PI diagnosis, making it urgent to develop such systems to improve diagnostic efficiency. MRI images, with their excellent tissue contrast and comprehensive information, are crucial for diagnosing PI. In this study, we developed a deep learning model based on multi-view MRI images. However, the scarcity of PI datasets limits model performance. To address this, we proposed a multi-task transfer learning approach that extracts shared feature layers by learning nine medically related indicators and applies transfer learning to enable the PI diagnostic model to acquire generalized feature extraction knowledge. The datasets were obtained from the Third Affiliated Hospital of Southern Medical University (538 cases) and Zhujiang Hospital (138 cases). The former was used for training, validation, and internal testing, while the latter was used for external testing. Our model achieved an AUC of 0.962 on the internal test set and 0.875 on the external test set, demonstrating strong generalization capability and significant potential for PI diagnosis.