Value of multi-label learning MRI model assisting radiological diagnosis of sports injury in knee / 中华放射学杂志

ABSTRACT

Objective:To construct a multi-label learning MRI model for assisting diagnosis of sports injury in knee.Methods:A total of 1 391 knee MRI cases from 1 343 young adults with sports injury in Affiliated Jinling Hospital Nanjing University School of Medicine were retrospectively enrolled. The image cases were randomly divided into training set ( n=973), validation set ( n=139) and test set ( n=279) with ratio of 7∶1∶2. The knee injuries were divided into six categories meniscus injury, tendon injury, ligament injury, osteochondral injury, synovial bursa disorder and soft tissue injury. Using PyTorch V1.1.0 algorithm package, the Yolo model of deep learning was used to construct the MRI knee joint sports injury detection model. The model was validated on the test set, and the sensitivity, specificity and mean average precision of lesion detection were evaluated. Results:Among the 279 patients in test set, the mean average precision of meniscus injury, tendon injury, ligament injury, osteochondral injury, synovial bursa disorder and soft tissue injury were 83.1%, 89.0%, 88.0%, 85.8%, 85.5% and 83.2%, respectively, and the overall mean average precision was 85.8%. The model was most effective in detecting tendon injury. The sensitivity and specificity of the model for tendon injury were 91.2% and 87.1% respectively.Conclusions:The multi-label MRI knee joint exercise-related injury detection model based on deep learning can effectively assist in detecting the exercise-related injury of knee joint in each tissue structure, and is expected to improve the efficiency of diagnosis and treatment in orthopedics.