Macro Habitat‐Based <scp>T2</scp> ‐Weighted <scp>MRI</scp> Radiomics and Deep Learning Fusion for Predicting Treatment Response and Prognosis After Neoadjuvant Chemoradiotherapy in Locally Advanced Rectal Cancer

BACKGROUND: Response to neoadjuvant chemoradiotherapy (NACRT) in locally advanced rectal cancer (LARC) is highly heterogeneous. Reliable pretreatment prediction of tumor regression and prognosis remains an unmet clinical need to optimize personalized management. METHODS: A multicenter retrospective study of 434 Stage II-III LARC patients was conducted across three tertiary hospitals. Pretreatment T2-weighted MRI was used to build intratumoral and peritumoral macrohabitats. Local radiomic features within the tumor were clustered using K-means to generate intratumoral habitats, with the optimal cluster number determined by the Calinski-Harabasz score. Radiomic and 3D deep-learning features from each habitat and the peritumoral region were fused after LASSO-based selection. Machine-learning classifiers (support vector machine, logistic regression, multilayer perceptron) were trained to predict tumor regression grade (TRG). Performance was assessed by ROC and decision-curve analyses, and prognostic value for progression-free survival (PFS) was evaluated using Kaplan-Meier analysis. RESULTS: The macro habitat-based fusion model demonstrated superior performance compared with intratumoral, peritumoral, or single-habitat models, achieving AUCs of 0.807-0.830 in the external validation cohort. The derived risk score showed a significant association with progression-free survival (PFS) (p = 0.011 in the training and p = 0.030 in the validation cohorts). CONCLUSIONS: The macro habitat-based MRI radiomics and deep learning fusion model provides a noninvasive, interpretable, and robust biomarker for predicting treatment response and prognosis in LARC. It holds potential to guide personalized therapeutic strategies, including organ-preserving approaches and tailored surveillance after NACRT.