Comprehensive Summary
Multicenter retrospective study of 560 stage II-III colorectal cancer patients from three hospitals (training n = 282; external tests n = 88 and n = 111) between July 2018 and March 2022; median follow-up 35 months with 105 early recurrences within two years. Venous-phase CT tumors were segmented by radiologists; 1,066 stable handcrafted features and 512 deep features were reduced to 5 radiomics plus 8 deep features to form a fused radiomics–deep learning signature (RDs). RDs predicted early recurrence with AUC 0.841 (training), 0.858 (test 1), and 0.835 (test 2). A preoperative combined model that included RDs and blood indices was independently associated with early recurrence: RDs OR 781.77 (95% CI 94.67-6455.96, p < 0.001), lymphocyte-to-monocyte ratio OR 0.30 (0.10-0.90, p = 0.032), neutrophil-to-lymphocyte ratio OR 3.61 (1.09-11.89, p = 0.035), lymphocyte-albumin OR 4.19 (1.01-17.49, p = 0.049), prognostic nutritional index OR 0.24 (0.08-0.72, p = 0.011). The combined model outperformed postoperative pathology and TNM in both external cohorts (AUC 0.865 and 0.848; accuracy 0.875 and 0.667; sensitivity 0.800 and 0.963). Using a cutoff of 0.102, low-risk patients had significantly longer disease-free survival across stage and chemotherapy strata (p ≤ 0.003), and an external GEO set (n = 79) showed 40 differentially expressed genes with enrichment of epithelial-mesenchymal transition and inflammatory pathways in high-risk biology.
Outcomes and Implications
Before surgery, clinicians can estimate a patient’s two-year recurrence risk using a score built from CT images and routine blood tests, not just stage. In two outside hospitals, this preoperative score identified high-risk patients more accurately than TNM stage or standard postoperative pathology, which means decisions can be made earlier. High-risk patients can be considered for neoadjuvant or intensified perioperative therapy, closer imaging follow-up, and targeted efforts to improve nutrition and inflammatory status. Low-risk patients can stay on standard care pathways and avoid extra treatment that is unlikely to help. The score also split outcomes within both stage II and stage III groups, with or without chemotherapy, so it adds useful information beyond usual staging. Findings from an external dataset point to epithelial–mesenchymal transition and inflammatory programs in high-risk tumors, which can guide trial enrollment and future therapy choices; prospective validation and a fixed imaging pipeline are needed before routine use.