Deep learning model based on DCE-MRI: fusion of 3D features of tumor, peritumoral vessels and metastatic lymph nodes for prediction of pathological complete response to neoadjuvant therapy in breast cancer.

Objective: The aim of this study is to develop a deep learning-based radiomic (DLR) model by fusing 3D features of tumor, peritumoral vessels, and metastatic lymph nodes from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), with the goal of predicting pathological complete response (pCR) in breast cancer patients receiving neoadjuvant therapy. Materials and methods: A total of 200 breast cancer (BC) cases were retrospectively collected from the First and Second Affiliated Hospitals of Bengbu Medical University between January 2020 and December 2024. The cases were randomly allocated to a training set and a test set at a 1:1 ratio. For dynamic contrast-enhanced MRI (DCE-MRI) sequence imaging, 3D UNet technology was utilized to facilitate layer-by-layer semi-automated segmentation of tumors, peritumoral vessels, and metastatic lymph nodes. Concurrently, we used deep learning methods to extract features and constructed a predictive model for pCR status in breast cancer patients after NAT. The Clinical Combined Deep Learning Radiomic (CCDLR) model was developed by integrating clinical characteristics into the DLR model. The performance of the CCDLR and DLR models was compared and validated in a test set. Results: The training set contained 45 cases in the pCR group and 55 cases in the non-pCR group, while the test set contained 47 cases in the pCR group and 53 cases in the non-pCR group. The efficacy of the CCDLR model in predicting the NAT pCR of breast cancer was superior to that of the DLR model. The AUC values of the CCDLR model and the DLR model in the training set were 0.950 and 0.820, with accuracies of 96.0% and 81.0%, precision of 95.1% and 79.6%, recall of 95.1% and 84.3%, and F1 scores of 95.1% and 81.9%.In the test set, the AUC values of the two models were 0.870 and 0.850, with accuracies of 92.0% and 83.0%, precision of 92.1% and 83.3%, recall of 92.1% and 73.1%, and F1 scores of 92.1% and 77.9%. Conclusion: Fusing three-dimensional imaging features of tumors, peritumoral vessels, and metastatic lymph nodes, the DLR model shows favorable predictive efficacy. Importantly, the CCDLR model, constructed by incorporating clinical characteristics, exhibits significantly superior performance, underscoring its promising potential for clinical application in predicting pathological complete response (pCR) to neoadjuvant therapy (NAT) in breast cancer.