MRI-based Diagnosis of Rotator Cuff Tears using Deep Learning and Weighted Linear Combinations
Proceedings of the 5th Machine Learning for Healthcare Conference, PMLR 126:292-308, 2020.
Rotator Cuff Tears (RCTs) are a common injury among people who are middle-aged or older. For effective diagnosis of RCTs, orthopedic surgeons typically need to have access to both shoulder Magnetic Resonance Imaging (MRI) and proton density-weighted imaging. However, the generation and interpretation of such comprehensive image information is labor intensive, and thus time consuming and costly. Although computer-aided diagnosis can help in mitigating the aforementioned issues, no computational tools are currently available for diagnosing RCTs. Therefore, we introduce a computational approach towards RCT diagnosis in this paper, leveraging end-to-end learning by applying a deep convolutional neural network to shoulder MRI scans. Given that these shoulder MRI scans are 3-D by nature and highly biased towards normal shoulders, with only 6.6% of the available shoulder MRI scans containing partial-thickness tears, we made use of two tools to enhance our deep convolutional neural network. First, to enable the utilization of sequential information available in the 3-D MRI scans, we integrated a weighted linear combination layer. Second, to mitigate the presence of class imbalance, we adopted weighted cross-entropy loss. That way, we were able to obtain a diagnostic accuracy of 87% and an M-AUC score of 97%, outperforming a baseline of human annotators (diagnostic accuracy of 76% and an M-AUC score of 81%). In addition, we were able to outperform several approaches using conventional machine learning techniques. Finally, to facilitate further research efforts and ease of benchmarking, we make our dataset of 2,447 shoulder MRI scans publicly available.