Integrating MRI habitat heterogeneity and peritumoral radiomics into a nomogram for optimized risk stratification in locally advanced rectal cancer: a multicenter study

Background: Risk stratification is essential for optimizing treatment in locally advanced rectal cancer (LARC), particularly for identifying suitable candidates for total neoadjuvant therapy (TNT). Conventional MRI-based staging has limited sensitivity in capturing tumor microenvironment (TME) heterogeneity, which may lead to undertreatment of high-risk patients or overtreatment of low-risk subgroups. This study aimed to develop a nomogram integrating MRI-based habitat heterogeneity and peritumoral radiomics to improve risk stratification in LARC. Methods: A multicenter retrospective cohort of 290 LARC patients (training set, n=178; external test set, n=112) was analyzed. Tumor volumes and peritumoral regions (1, 2, and 3 mm margins) were delineated on high-resolution MRI scans using 3D Slicer. Habitat heterogeneity was quantified via K-means clustering (k=3) of intratumoral radiomic features. Radiomic features were filtered and reduced using LASSO regression. Logistic regression (LR) and support vector machine (SVM) classifiers were used to build intratumoral, peritumoral, and habitat models. The better-performing model between the intratumoral and habitat models was combined with the optimal peritumoral model and clinical variables to construct a nomogram. Calibration curves assessed agreement between predicted and observed high-risk LARC. Model performance was evaluated using area under the curve (AUC), sensitivity, specificity, and decision curve analysis (DCA). Results: LR classifiers outperformed SVM classifiers and were therefore selected for the intratumoral, peritumoral, and habitat models. The habitat and peritumoral (3 mm) models showed superior performance compared with the intratumoral and peritumoral (1 mm) models and were integrated with clinical variables into a nomogram. The nomogram achieved excellent performance in both training (AUC, 0.928) and test cohorts (AUC, 0.817), surpassing single-feature models. Calibration curves demonstrated good agreement between predicted and observed high-risk LARC. DCA showed the nomogram provided higher net benefit across a broad range of threshold probabilities. Conclusion: By characterizing the spatial heterogeneity of the tumor microenvironment, an MRI-derived nomogram integrating habitat heterogeneity, peritumoral (3 mm) radiomic features, and clinical variables was developed to facilitate precise risk stratification and personalized TNT decision-making for LARC.