Deep-learning-assisted diagnosis for knee magnetic resonance imaging: Development and retrospective validation of MRNet
Fig 3
Class activation mappings for MRNet interpretation.
Class activation mappings (CAMs) highlight which pixels in the images are important for the model’s classification decision. One of the board-certified musculoskeletal radiologists annotated the images (white arrows and circles) and provided the following captions. (a) Sagittal T2-weighted image of the knee demonstrating large effusion (arrow) and rupture of the gastrocnemius tendon (ring), which were correctly localized by the model and classified as abnormal. Note that the model was not specifically trained to detect these pathologies but was able to recognize the abnormalities based on the contrast with the normal knee examinations. (b) Sagittal T2-weighted image of the knee complicated by a significant motion artifact demonstrating complete anterior cruciate ligament (ACL) tear (arrow), which was correctly classified and localized by the model. Because we hoped to best approximate the clinical practice reality—in which the prevalence of artifacts (i.e. motion, metallic) and other technical noise disrupts interpretation of knee MRI—we did not exclude noisy cases from the training or validation data. (c) Sagittal T2-weighted image of the knee demonstrating complete disruption of the ACL, which was correctly identified by the model as abnormal and classified as ACL tear. The CAM indicates the focus of the model at the abnormal attachment of the ACL (arrow). (d) Sagittal T2-weighted image of the knee demonstrating a complex tear involving the posterior horn of the lateral meniscus (arrow). While the model did classify this examination as abnormal, the CAM indicates that the increased subcutaneous signal (ring) in the anterior/lateral soft tissues contributed to the decision but the meniscal tear did not.
