Deep learning-driven MRI radiomics reveals biological subtypes and predicts recurrence risk in rectal cancer

Traditional TNM staging inadequately captures the recurrence risk of rectal cancer (RC), limiting prognostic accuracy and personalized treatment decisions. Here, we developed an MRI-based deep learning approach to predict recurrence risk and leveraged pathology and transcriptomic analyses to characterize tumor heterogeneity and provide biological interpretation. In this multicenter study of 2060 patients without neoadjuvant therapy across four independent cohorts (training: n = 931; internal validation: n = 365; external validation: n = 492; biological: n = 272), unsupervised clustering of deep learning-extracted radiomic features identified three biologically distinct radiomics-based deep learning subtypes (RDLSs). Comprehensive molecular profiling using single-cell RNA sequencing (n = 7), bulk RNA sequencing (n = 83), and whole-slide imaging (n = 409) revealed that RDLS1 exhibited an immune-excluded phenotype with poor prognosis, RDLS2 showed enriched lymphocyte infiltration with favorable outcomes, and RDLS3 displayed intermediate prognosis with abundant stromal elements. Multivariate Cox analysis confirmed independent prognostic value across all cohorts (RDLS2 vs. RDLS1: HR = 0.52, 95% CI: 0.29-0.82, p = 0.003). Leveraging high-risk RDLS1 features, we developed a recurrence risk scoring system (RRS) that effectively stratified patients across cohorts. Compared with the clinical model alone, integration of the RRS with clinical variables improved 5-year RFS prediction performance, increasing the AUC from 0.799 to 0.834 in the training cohort and from 0.780 to 0.825 in the external validation cohort. This noninvasive and biologically interpretable framework bridges imaging phenotypes to molecular characteristics, providing a potential approach for recurrence risk assessment and more individualized postoperative risk stratification in RC patients without neoadjuvant therapy.