Anomaly Detection for Knee Cartilage MRlmages Using Deep Transfer Learning Models

The potential impact of undetected (asymptomatic) knee ligament tears in healthy individuals remains a significant concern. This study aimed to investigate the prevalence of potential knee abnormalities using magnetic resonance imaging (MRI) data. The retrospectively analyzed 1,370 knee MRI samples collected at Stanford University Medical Center. The data included 1,104 exams (80.6%) with identified abnormalities, such as meniscal tears (508, 37.1 %) and ACL tears (319, 23.3%). These findings were manually labeled based on clinical reports. The method employed is progressive deep transfer learning models, including Xception, ResNet50, VGG16, InceptionV3, and EfficientNet-B0, to classify ACL tears, abnormal scans, and meniscal tears across three MRI views (axial, coronal, sagittal). The performance of the EfficientNet-B0 is in good contrast to the other transfer learning models, with an average accuracy of 0.836 for Abnormal, 0.673 for ACL, and 0.666 for Meniscus. These findings provide valuable insights into the prevalence of knee abnormalities and the efficacy of deep transfer learning techniques in categorizing knee abnormalities from MRI data. This knowledge has the potential to significantly impact the field of radiology, enhancing our ability to diagnose and treat knee abnormalities.