A two-stage deep-learning model using MobileNetV2 and U-Net for CT-based muscle volume assessment in total hip arthroplasty

Preoperative muscle strength is an important predictor of functional recovery after total hip arthroplasty (THA); however, accurate assessment is difficult in patients with severe hip pain. In this study, we developed a two-stage deep-learning framework that combines slice-level classification and U-shaped convolutional neural network-based segmentation to automatically quantify periarticular and thigh muscle volumes from preoperative computed tomography (CT) images. The model was trained using CT scans from 107 patients undergoing primary THA and was validated against manual segmentation. The model was subsequently applied to an independent cohort of 58 patients to examine the associations between preoperative muscle volume indices and postoperative clinical outcomes. Automated segmentation demonstrated high agreement with manual measurements, with absolute volume errors of 1-10% and Dice coefficients exceeding 0.89 across all muscle groups. In the clinical cohort, the preoperative muscle volumes were significantly lower on the affected side than on the unaffected side. Greater preoperative volumes of the gluteus maximus, quadriceps, and hamstrings were significantly associated with postoperative hip abductor strength and Timed Up and Go test performance, whereas gluteus medius/minimus volumes showed weaker associations. These findings support the feasibility of automated CT-based muscle volumetry for objective functional assessment in THA.