Multimodal Fusion of Radiomics, Two-Dimensional Deep Learning, and Clinical Risk Factors for Preoperative Prediction of Pathological Complete Response to Neoadjuvant Chemotherapy in Breast Cancer: A Multicenter External Validation Study

Purpose Accurate preoperative prediction of pathological complete response (pCR) following neoadjuvant chemotherapy (NAC) is essential for individualized treatment decision-making in patients with breast cancer. This study aimed to develop a multimodal fusion model integrating conventional radiomics, two-dimensional (2D) deep learning features, and clinical risk factors based on the pre-treatment third-phase images of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), and to validate its efficacy for predicting pCR in an independent institutional cohort. Methods This retrospective multicenter study enrolled 372 women with breast cancer who received NAC across two centers. The 245 patients from Center 1 constituted the model development cohort and were stratified by pCR status and randomly split at a 7:3 ratio into a training set (n = 171) and an internal validation set (n = 74), whereas the 127 patients from Center 2 served as an independent external test set. A radiologist manually delineated the tumor region of interest on the third contrast-enhanced phase of DCE-MRI, and all delineations were reviewed by a senior radiologist. A total of 1,197 radiomic features were extracted using PyRadiomics; after Spearman correlation-based redundancy removal and dimensionality reduction with least absolute shrinkage and selection operator (LASSO) regression, 10 features were retained to construct the radiomics model. Six ImageNet-pretrained convolutional neural networks were compared, and MobileNet V2 was selected to build the 2D deep learning model. After dimensionality reduction by principal component analysis, the deep features were early-fused with the radiomic features to develop a deep learning radiomics (DLR) model. Independent clinical predictors of pCR were identified using univariate and multivariate logistic regression and used to construct a clinical model. The predicted probability of the DLR model was further fused with the independent clinical factors using logistic regression to build the combined model, which was presented as a nomogram. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, calibration curves, decision curve analysis, the DeLong test, the integrated discrimination improvement (IDI), and the net reclassification improvement (NRI). Interpretability analyses were conducted using SHAP and Grad-CAM. Results The pCR rate of the entire cohort was 14.2% (53/372), with rates of 17.0%, 21.6%, and 6.3% in the training, internal validation, and external test sets, respectively. Multivariate logistic regression identified ER status (adjusted OR = 0.119, P < 0.001), HER-2 status (adjusted OR = 4.870, P = 0.003), and maximum tumor diameter (adjusted OR = 0.742, P = 0.041) as independent predictors of pCR. In the independent external test set, the combined model achieved the best overall discriminative performance, with an AUC of 0.933 (95% CI: 0.879–0.987), accuracy of 0.843, sensitivity of 1.000, specificity of 0.832, and negative predictive value of 1.000, outperforming the clinical model (AUC = 0.925), the deep learning model (0.825), the DLR model (0.745), and the radiomics model (0.559); its AUC was significantly higher than those of the radiomics model (P = 0.0007) and the DLR model (P = 0.014). The deep learning model showed the smallest cross-center performance degradation (ΔAUC = 0.033), markedly smaller than that of the conventional radiomics model (ΔAUC = 0.278). Relative to the four single-modality models, the combined model yielded positive IDI (0.208–0.482, all P < 0.05) and NRI (0.216–0.683) values, good calibration (Hosmer–Lemeshow test, P = 0.069), and a higher net clinical benefit on decision curve analysis. SHAP and Grad-CAM analyses consistently demonstrated that model decisions focused on intratumoral enhancement heterogeneity and peritumoral regions, with the most contributory features derived from both radiomics and deep learning, indicating complementarity between the two sources of information. Conclusions The multimodal fusion nomogram integrating independent clinical predictors, conventional radiomic features, and 2D deep learning features can accurately and robustly predict pCR after NAC in breast cancer before surgery, with favorable cross-center generalizability and interpretability. It may serve as a noninvasive quantitative tool for evaluating NAC response and supporting individualized treatment decision-making.