MRI radiomics-based predictive modeling for risk stratification and prognostication in oral tongue carcinoma

Abstract

Introduction This study aims to predict recurrence-free survival (RFS) and overall survival (OS), to stratify risk using radiomics, radiology semantic, clinical, and pathology models—individually and in combination—derived from pre-treatment magnetic resonance imaging (MRI), post-operative histopathology, age, and adjuvant therapy details in oral tongue squamous cell carcinoma (OTSCC), and to correlate National Comprehensive Cancer Network (NCCN) low + intermediate- and high-risk groups with the risk stratification provided by the radiomics models. The secondary objectives included predicting post-operative perineural invasion (PNI), metastatic lymph nodes, and pathological extranodal extension (pENE). Methods This retrospective cohort study included 219 treatment-naive OTSCC patients who underwent pre-treatment MRI (January 2021–December 2022) with 2-year follow-up. Radiomics features were extracted from T1-weighted, T2-weighted, and contrast-enhanced T1-weighted (T1c) sequences. Lasso–Cox and DeepSurv models were evaluated for RFS and OS prediction, while random forest and logistic regression models were assessed for PNI, metastatic lymph nodes, and pENE prediction. Radiomics, radiology semantic, clinical, and pathology models were evaluated individually and in combination. Risk stratification was evaluated using the log-rank test. Results Among 219 patients (mean age 46.99 ± 10.94 years), Lasso–Cox outperformed DeepSurv for RFS and OS prediction. The radiomics combined model (T1 + T2 + T1c radiomics + radiology semantic + clinical + pathology) achieved the best performance for RFS and non-inferior performance to the pathology model for OS prediction. The pre-treatment radiomics models effectively stratified the patients into high- and low-risk groups for recurrence and OS and also identified low-risk patients within the NCCN high-risk group and high-risk patients within the NCCN low + intermediate-risk group for recurrence. The T1c radiomics model demonstrated the best overall performance for predicting metastatic lymph nodes. The radiology semantic features model demonstrated the highest predictive performance for PNI and pENE. Conclusions The radiomics combined and pre-treatment radiomics models demonstrate potential for prognostication and risk-adapted management. By identifying low-risk patients within the NCCN high-risk group and high-risk patients within the NCCN low + intermediate-risk group for recurrence, these models may support treatment de-escalation and escalation, respectively, pending prospective multicenter validation. Furthermore, prediction of metastatic lymph nodes, PNI, and pENE from pre-treatment radiomics models may inform decisions regarding elective neck dissection and guide adjuvant therapy following prospective validation.