Application of a MRI model based on tumor microenvironment habitat and peritumoral features in preoperative differentiation of rectal cancer T1-2/T3: a multicenter study

Aim To develop a MRI model integrating tumor microenvironment (TME)-driven habitat analysis, peritumoral radiomics, and clinical risk factors for improving preoperative differentiation of T1-2 and T3 stages in rectal cancer. Methods This retrospective multicenter study included 313 patients (training cohort: 183; external test cohort: 130). MRI-derived intratumoral radiomic features were clustered into three TME habitats using K-means, while peritumoral features (1 mm, 2 mm, 3 mm extensions) were extracted. Feature selection and model construction were performed using LASSO regression and cross-validation, with all steps strictly confined to the training cohort. Four models were evaluated: peritumoral (PERI2mm optimal), habitat, clinical, and a nomogram combining radiomic features (habitat and PERI2mm) with clinical predictors. Performance was assessed using ROC curves, calibration metrics (intercept, slope, Brier score), and decision curve analysis (DCA). Results The nomogram achieved the highest AUC of 0.907 (training) and 0.881 (test), outperforming standalone models in the test cohort (PERI2mm: AUC = 0.721; habitat: AUC = 0.741; clinical: AUC = 0.723). The nomogram also demonstrated superior balanced accuracy (0.827), PR-AUC (0.948), and F1 score (0.851) in the test cohort. Key clinical predictors included age, elevated CEA, tumor length, circumferential growth and mrT stage. Calibration analysis confirmed excellent agreement between predicted and pathological staging, with the nomogram showing the lowest Brier scores (0.1214 training, 0.1253 test) and intercepts near zero, indicating no significant systematic bias. DCA demonstrated superior clinical net benefit across risk thresholds. Conclusion The TME-integrated nomogram significantly enhances preoperative T1-2/T3 staging accuracy in rectal cancer by leveraging habitat heterogeneity, peritumoral radiomics, and clinical biomarkers, with robust calibration and generalization. This tool may refine therapeutic strategies, reduce overtreatment, and improve patient outcomes.