A longitudinal CT-based subregional radiomics nomogram for predicting local recurrence-free survival in esophageal squamous cell carcinoma after definitive chemoradiotherapy: a multicenter study.

Abstract

Accurate prediction of local recurrence-free survival (LRFS) is crucial for personalizing treatment in esophageal squamous cell carcinoma (ESCC) patients undergoing definitive chemoradiotherapy (dCRT). This multicenter study enrolled 306 ESCC patients from two institutions. A subregional radiomics workflow was applied to contrast-enhanced CT scans acquired before and after dCRT to extract radiomic features from intratumoral and peritumoral subregions and construct a robust fusion signature (Subrad) based on generalized boosted regression model (GBM). A clinic-radiomics nomogram (Clirad) was subsequently developed by integrating the Subrad signature with clinical factors. The interpretability of the model is enhanced through Shapley additive explanations (SHAP) and radiogenomics analysis. The Subrad signature significantly outperformed the clinical model (C-index: 0.706 vs. 0.629, P = 0.016) and a conventional pretreatment radiomics signature (Prerad) (C-index: 0.598) in the external testing set. The final Clirad nomogram, incorporating the Subrad signature and concurrent chemotherapy, achieved the highest prognostic value, with C-indices of 0.896, 0.725, and 0.708 in the training, internal testing, and external testing sets, respectively. It effectively stratified patients into distinct high- and low-risk groups for LRFS (P < 0.05). Additionally, RNA sequencing analysis provided a preliminary exploration of the biological relationships between the tumor microenvironment and the Subrad signature. The Clirad nomogram provides a valuable tool for predicting LRFS, offering an understanding of tumor spatial heterogeneity and supporting personalized treatment planning in EC.