A Nomogram Integrating Deep Learning and Immunoscore for Predicting Distant Metastasis and Prognosis in Rectal Cancer

To evaluate the performance of nomogram integrating MRI-based deep learning (DL) and immunoscore (IS) for predicting distant metastasis (DM) and prognostic stratification in rectal cancer (RC).

This retrospective single-center study included 306 RC patients (mean age, 59.7±11.2 years) who underwent MRI and surgery. A DL model was developed from preoperative T2-weighted images. The IS was determined on surgical specimens via immunohistochemistry for CD3+ and CD8+ T cells. A nomogram was constructed by integrating the DL_score, IS, and clinical predictors identified by Cox regression. Model performance was assessed using the concordance index (C-index), area under the curve (AUC), calibration curves, and decision curve analysis. Feature importance was measured by the SHapley Additive exPlanations method and patient stratification was assessed through Kaplan-Meier analysis.

At the cutoff date, 97 (31.7%) had experienced DM or death, with a median follow-up of 45.8 months (IQR, 33.9-85.8). Patients with lower IS showed significantly higher DM risk and worse DM-free survival (P < 0.001). The DL_score, IS, and clinical predictors were used to construct the nomogram with C-indexes of 0.807 and 0.882 in training and test set, with AUCs of 0.922 and 0.914 for predicting 5-year DM-free survival. Nomogram effectively stratified patients into high- and low-risk groups and high-risk subgroups had significantly worse DM-free survival (Log-rank, P < 0.001).

The integrated nomogram demonstrates strong performance for individualized DM risk prediction and prognostic stratification in RC, providing a clinically useful tool for postoperative management.