Abstract Objectives To develop and validate a deep learning model for automated masseter muscle segmentation on routine head and neck CT and to evaluate whether the derived masseter muscle area is associated with overall survival in oral cancer. Materials and methods A U-Net-based model was trained using a model-development cohort (n = 348) with preoperative CT and bioelectrical impedance analysis. Masseter muscle area was measured on an axial slice at the maxillary sinus floor, and sex-specific cutoffs for low masseter muscle area were derived against sarcopenia defined by appendicular skeletal muscle index. Prognostic value of AI-derived masseter muscle area (AI-MMA) was tested in an independent cohort of primary oral cancer patients (n = 247) using Kaplan–Meier analysis and Cox proportional hazards models. Results Segmentation performance was high (Dice similarity coefficient, 0.92). AI-MMA correlated strongly with manual MMA in males and females (r = 0.892 and r = 0.896, respectively; both p < 0.001). Low AI-MMA was associated with poorer overall survival. In multivariable analysis, low AI-MMA remained an independent predictor of mortality (hazard ratio HR, 2.584; 95% CI, 1.132–5.898; p = 0.024), together with stage III–IV disease (HR, 5.811; 95% CI, 2.130–15.860; p < 0.001) and low body mass index (HR, 2.572; 95% CI, 1.162–5.693; p = 0.020). Conclusions Automated AI-MMA from routine staging CT provides an objective prognostic biomarker in oral cancer.
Hiraoka et al. (Sat,) studied this question.