CT-based radiomics signature for optimizing adjuvant chemotherapy decisions in stage II colorectal cancer with proficient mismatch repair.

[PURPOSE] To develop a CT-based radiomics model for predicting prognosis and optimizing adjuvant chemotherapy (ACT) decisions in stage II proficient mismatch repair (pMMR) colorectal cancer (CRC).

[METHODS] We retrospectively enrolled 379 stage II pMMR CRC patients with distinctly good or poor outcomes, including 125 patients in the validation cohort. The radiomics model was developed using features extracted from segmented CT volumes by XGBoost. Clinical model A was constructed via multivariate logistic regression incorporating clinical factors recommended by the European Society of Medical Oncology (ESMO). Model performance was compared using the DeLong test. Survival analysis employed Kaplan-Meier curves and log-rank tests, and ACT subgroup analysis was performed according to risk stratification.

[RESULTS] In the validation cohort, the radiomics model (AUC = 0.800) and combined model A (AUC = 0.811) outperformed clinical model A (AUC = 0.576; both p < 0.05). Patients stratified as low-risk by either the radiomics model or the combined model A exhibited significantly longer overall survival (OS) than high-risk patients in both training and validation cohorts (all p < 0.05). Within the high-risk groups identified by the radiomics model and the combined model A, patients receiving ACT showed significantly longer OS compared to non-recipients (p = 0.015 and p = 0.008, respectively), while no significant difference was observed in low-risk groups (both p > 0.05).

[CONCLUSION] Compared to ESMO-recommended clinical high-risk factors, the CT-based radiomics model demonstrated superior prognostic prediction performance and could optimize ACT decisions based on risk stratification in stage II pMMR CRC.