Abstract
OBJECTIVE: The aim of this study was to develop a multimodal fusion model for accurate risk prediction and clinical decision support for ductal carcinoma in-situ (DCIS). METHOD: By integrating deep learning (DL), radiomics and clinical features, this study constructed the combined model and validated its performance in a multicenter cohort containing 232 patients (103 in the training set, 43 in the validation set and 86 in the external test set). RESULTS: Unimodal DL models showed significant overfitting in external tests (e.g., DenseNet201 training set AUC = 0.85 vs. test set 1 AUC = 0.47), whereas the multimodal fusion model achieved optimal predictive performance across cohorts through heterogeneous data synergy (training set AUC = 0.925, test set 2 AUC = 0.801) with the DeLong test confirmed that it significantly outperformed the unimodal model (P < 0.05). Grad-CAM visualization showed that the model focus region was highly consistent with the radiologist annotation (81% overlap, Cohen's κ = 0.68). Calibration curves (Hosmer-Leeshawn test P > 0.05) and decision curve analysis (DCA) validated the model prediction reliability (error < 5%) with a net clinical benefit advantage (net benefit difference of 7% to 28% at thresholds of 5% to 80%). CONCLUSION: The multimodal fusion strategy could potentially mitigate the limitations of unimodal models, presenting a potential solution with promising high accuracy and interpretability for individualized diagnosis and treatment of DCIS.