Deep Learning-Based Muscle Segmentation and Quantification of Full-Leg Plain Radiograph for Sarcopenia Screening in Patients Undergoing Total Knee Arthroplasty

Abstract

Sarcopenia, an age-related loss of skeletal muscle mass and function, is correlated with adverse outcomes after some surgeries. Here, we present a deep-learning-based model for automatic muscle segmentation and quantification of full-leg plain radiographs. We illustrated the potential of the model to predict sarcopenia in patients undergoing total knee arthroplasty (TKA). A U-Net-based deep learning model for automatic muscle segmentation was developed, trained and validated on the plain radiographs of 227 healthy volunteers. The radiographs of 403 patients scheduled for primary TKA were reviewed to test the developed model and explore its potential to predict sarcopenia. The proposed deep learning model achieved mean IoU values of 0.959 (95% CI 0.959-0.960) and 0.926 (95% CI 0.920-0.931) in the training set and test set, respectively. The fivefold AUC value of the sarcopenia classification model was 0.988 (95% CI 0.986-0.989). Of seven key predictors included in the model, the predicted muscle volume (PMV) was the most important of these features in the decision process. In the preoperative clinical setting, wherein laboratory tests and radiographic imaging are available, the proposed deep-learning-based model can be used to screen for sarcopenia in patients with knee osteoarthritis undergoing TKA with high sarcopenia screening performance.