Improved risk stratification for oral tongue squamous cell carcinoma using deep learning-based MRI classifier and pathological markers to guide postoperative radiotherapy: A multicenter cohort study.

Abstract

Current risk stratification for oral tongue squamous cell carcinoma (OTSCC) has limited accuracy in identifying patients who would benefit from postoperative radiotherapy (PORT), resulting in potential over- or under-treatment. We aimed to develop a more precise risk stratification system for OTSCC by integrating deep learning (DL)-based imaging analysis with pathological markers. A DL model was developed utilizing preoperative MRI images from 658 patients, and externally validated in 419 patients from two independent centers. Subsequently, we constructed a pathological deep learning nomogram (PDLN) that integrated the DL-based risk score with pathological markers (pT stage, pN stage, tumor grade, and tumor budding). Model performance was assessed using the C-index and time-dependent ROC curve analysis. Concomitantly, a composite score combining the DL model and pathological markers was established, forming the composite risk integration system for OTSCC (CRIS-OTSCC). Furthermore, we also evaluated the prognostic accuracy of the CRIS-OTSCC as well as its association with PORT. The DL model showed a consistently high accuracy in predicting outcomes across all cohorts (C-index > 0.700). In multivariable analysis, after adjusting for clinicopathological variables, the DL-based risk score independently predicted both OS (overall survival) and DFS (disease-free survival) in all cohorts (all P < 0.001). The PDLN further improved the performance, with a C-index of 0.850 (95% CI, 0.815-0.885) for OS and 0.763 (0.718-0.808) for DFS in the training cohort, 0.809 (0.750-0.868) and 0.726 (0.655-0.797) in the internal validation cohort, 0.804 (0.747-0.861) and 0.737 (0.680-0.794) in the external validation cohort-A, and 0.778 (0.707-0.849) and 0.736 (0.663-0.809) in the external validation cohort-B, respectively. Notably, for patients in the CRIS-OTSCC favorable group, PORT did not affect OS (hazard ratio [HR] 0·429 [0·111-1·660]; P = 0.220) and DFS (0·899 [0·449-1·801]; P = 0.764). Conversely, for patients in the CRIS-OTSCC poor group, PORT was associated with improved OS (0·194 [0·115-0·325]; P < 0.001) and DFS (0·332 [0·214-0·513]; P < 0.001). The CRIS-OTSCC not only allows accurate prediction of patient outcomes but also identifies high-risk patients who may benefit from PORT, whereas its potential to guide PORT sparing in low-risk patients requires further validation.