Radiomics-Driven Imaging Biomarkers for Predicting Chemoradiotherapy Response in Locally Advanced Rectal Cancer.

PURPOSE: Pathological complete response (pCR) after neoadjuvant chemoradiotherapy (nCRT) in locally advanced rectal cancer (LARC) predicts excellent prognosis, yet conventional morphological MRI achieves only 52-71% sensitivity. Radiomics models show promising discrimination, but performance variation across platforms limits clinical translation. METHODS: This prospective cohort study enrolled 615 LARC patients (cT3-T4 or cN+) across four centers. Centers 1-3 contributed the development cohort (n = 405); Center 4 was the sole external validation site, contributing mutually exclusive 1.5T (n = 98; 19 pCR, 19.4%) and 3T (n = 112; 24 pCR, 21.4%) cohorts. Baseline MRI yielded 107 radiomics features (42 retained). Clinical-only, radiomics-only and combined models predicting pCR (ypT0N0) were fitted by LASSO logistic regression in Python, with ComBat harmonization estimated on development data only, DeLong testing, calibration assessment and decision curve analysis. RESULTS: Development pCR prevalence was 20.0% (81/405). The combined model achieved AUC 0.887 (95% CI 0.846-0.928; sensitivity 87.7% [71/81], specificity 84.6% [274/324]) versus radiomics-only 0.842 and clinical-only 0.723 (DeLong p < 0.001). External validation gave AUC 0.821 at 1.5T and 0.908 at 3T. ComBat improved 1.5T discrimination (ΔAUC + 0.060, p = 0.041) but did not abolish the field-strength gradient. Calibration slopes were 0.91-0.96. At a 15% threshold probability the combined model showed the highest net benefit (0.135). pCR predicted 3-year disease-free survival 98.8% versus 66.0% (p < 0.001). CONCLUSIONS: Combined radiomics-clinical models discriminate pCR across MRI platforms. Because external validation was confined to one center and no management decision followed model output, these findings are hypothesis-generating and require prospective interventional verification.