ObjectiveTo develop a multimodal prediction model integrating radiomics, deep learning (DL) and nomogram, and evaluate its performance in predicting pathologic complete response (pCR) after total neoadjuvant therapy (TNT) for locally advanced rectal cancer (LARC). MethodsPatients with LARC who underwent radical resection after TNT from January 2023 to December 2024 were prospectively enrolled and randomly divided into a training cohort and a test cohort at a ratio of 7∶3. In the training cohort, radiomic features were extracted from baseline T2-weighted imaging and diffusion-weighted imaging; The least absolute shrinkage and selection operator regression was adopted to screen the optimal feature subsets, based on which a radiomics score formula was constructed, and the radiomics score of each patient was subsequently calculated; A dual-channel DL model based on ResNet-50 architecture was constructed simultaneously; Univariate and multivariate logistic regression analyses were performed to screen independent predictive factors for the development of a nomogram model. A stacking ensemble strategy was adopted to integrate the above three base models into a final fusion model. The area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, Brier score, Hosmer-Lemeshow test, integrated discrimination improvement (IDI) index, net reclassification improvement (NRI) index, decision curve analysis (DCA), and subgroup analysis were applied to comprehensively evaluate the performance and clinical value of all models. ResultsA total of 196 patients were enrolled, among whom 58 patients achieved pCR, accounting for 29.6%. Among 196, 137 cases were assigned to the training cohort and 59 cases to the test cohort. Based on the training cohort data, five independent predictive factors were identified using multivariate logistic regression to develop a nomogram, including tumor distance from the anal verge, baseline carcinoembryonic antigen (CEA), neutrophil-to-lymphocyte ratio, ≥60% reduction in CEA at mid-chemotherapy (after 2 cycles), and interval from the end of radiotherapy to surgery. The DL model converged after 50 training epochs, radiomics score [M（P25, P75）] was 0.45 (–0.32, 1.58). A stacking ensemble fusion model was established by integrating three base models, namely the radiomics model, DL model, and nomogram. In the test cohort, the AUC of the fusion model was 0.893, which was higher than 0.801 for the radiomics model, 0.841 for the DL model, and 0.788 for the nomogram. All differences were statistically significant by the DeLong test (all P<0.05). The sensitivity, specificity, accuracy, and Brier score of the fusion model were 88.24%, 83.33%, 84.75%, and 0.112, respectively. The Hosmer-Lemeshow test suggested a good consistency between the predicted and actual probabilities (P=0.522). The IDI index of the fusion model vs. the DL model and vs. the nomogram was 0.128 and 0.156, respectively; the NRI index was 0.342 and 0.287 (all P<0.001), respectively. DCA demonstrated that the fusion model yielded the highest clinical net benefit within the threshold range of 10% to 60%. Subgroup analysis showed that the fusion model maintained favorable predictive performance across subgroups stratified by tumor location, clinical stage, baseline CEA level, age, gender, and interval from the end of radiotherapy to surgery, with AUC values ranging from 0.856 to 0.918. Particularly, the AUC reached 0.918 in patients with tumor distance from the anal verge <5 cm and 0.911 in those with interval from the end of radiotherapy to surgery ＞10 weeks. ConclusionsThe fusion model constructed by the stacking ensemble strategy integrates multi-phase magnetic resonance imaging radiomic features, dynamic biomarker changes, and clinicopathologic parameters. It exhibits excellent discriminative ability, calibration performance and clinical practicability, and maintains robust stability across different clinical subgroups. This model can accurately predict the status of pCR in LARC patients after TNT. Compared with single models, the fusion model achieves better predictive efficacy and higher clinical net benefit, and can provide references for individualized clinical treatment decisions.