Evaluation of a deep learning method for the automated detection of supraspinatus tears on MRI

To evaluate if deep learning is a feasible approach for automated detection of supraspinatus tears on MRI. A total of 200 shoulder MRI studies performed between 2015 and 2019 were retrospectively obtained from our institutional database using a balanced random sampling of studies containing a full-thickness tear, partial-thickness tear, or intact supraspinatus tendon. A 3-stage pipeline was developed comprised of a slice selection network based on a pre-trained residual neural network (ResNet); a segmentation network based on an encoder-decoder network (U-Net); and a custom multi-input convolutional neural network (CNN) classifier. Binary reference labels were created following review of radiologist reports and images by a radiology fellow and consensus validation by two musculoskeletal radiologists. Twenty percent of the data was reserved as a holdout test set with the remaining 80% used for training and optimization under a fivefold cross-validation strategy. Classification and segmentation accuracy were evaluated using area under the receiver operating characteristic curve (AUROC) and Dice similarity coefficient, respectively. Baseline characteristics in correctly versus incorrectly classified cases were compared using independent sample t-test and chi-squared. Test sensitivity and specificity of the classifier at the optimal Youden's index were 85.0% (95% CI: 62.1-96.8%) and 85.0% (95% CI: 62.1-96.8%), respectively. AUROC was 0.943 (95% CI: 0.820-0.991). Dice segmentation accuracy was 0.814 (95% CI: 0.805-0.826). There was no significant difference in AUROC between 1.5 T and 3.0 T studies. Sub-analysis showed superior sensitivity on full-thickness (100%) versus partial-thickness (72.5%) subgroups. Deep learning is a feasible approach to detect supraspinatus tears on MRI.