Multiparametric MRI-based Habitat Analysis Integrating Deep Learning and Radiomics for Predicting Preoperative Ki-67 Expression Level in Breast Cancer

Background Breast cancer (BC) is the most common malignant tumor in women globally. Ki-67, vital for prognosis, is currently detected invasively. Non-invasive MRI prediction faces challenges due to intratumoral heterogeneity. Materials and Methods This retrospective study included 254 breast cancer patients from two centers, divided into training (142 patients), internal validation set (60 patients), and external validation set (52 patients). T2WI and DCE-MRI sequences were analyzed. Traditional radiomics features were extracted from intratumoral, peritumoral (5 mm, 10 mm), and habitat regions. A pre-trained ResNet-50 model extracted 2.5D deep learning features. Feature selection used ICC, Z-score normalization, t-tests, Pearson correlations, and LASSO. Models were built using SVM, RF, and ET algorithms, evaluated via AUC, accuracy, sensitivity, specificity, and F1 score. SHAP analysis enhanced interpretability. Results The best-performing traditional radiomics model achieved an AUC of 0.825 (95% CI: 0.708–0.942) in the internal validation set. The optimal deep learning model obtained an AUC of 0.804 (95% CI: 0.641–0.966) in the internal validation set. The combined model, utilizing both best traditional radiomics and deep learning features, demonstrated superior performance with an AUC of 0.885 (95% CI: 0.787–0.984) in the internal validation set and 0.839 (95% CI: 0.727–0.951) in the external validation set. Conclusion The integrated model combining traditional radiomics and deep learning from MRI significantly predicts Ki-67 expression in breast cancer, enhancing preoperative prediction accuracy and interpretability for personalized treatment.