A deep learning radiomics model based on multiparametric MRI for predicting the efficacy of neoadjuvant chemotherapy in axillary lymph nodes of breast cancer.

OBJECTIVES: In this study, we developed and validated a deep learning radiomics clinical (DLRC) model integrating multiparametric MRI (mpMRI) and clinical data to predict axillary pathological complete response (apCR) in cN+ breast cancer patients receiving neoadjuvant chemotherapy (NAC). MATERIALS AND METHODS: We retrospectively collected data from 221 eligible breast cancer patients. The DLRC model integrated radiomics features and deep learning scores (DL-Scores) derived from mpMRI, together with clinical risk factors, to predict apCR. Seven classifiers were employed for model construction, and the optimal classifier was selected based on classification performance to build the final model. Model performance was evaluated using the receiver operating characteristic (ROC) curve, area under the curve (AUC), calibration curve, and decision curve analysis (DCA). Classification performance was assessed via the confusion matrix. Additionally, SHAP plots were utilized for interpretability analysis of the model. RESULTS: Multivariate analysis confirmed that progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), clinical T stage (OR = 0.258), and time-intensity curve (TIC) were significantly associated with apCR. The mpMRI based DLRC model achieved AUCs of 0.949 (95% CI, 0.912-0.986) and 0.932 (95% CI, 0.867-0.998) in the training and validation sets. SHAP analysis further revealed the interpretability of the model. The confusion matrix visually demonstrated the consistency between the predicted results and the true labels, validating the classification performance of the model. CONCLUSION: The DLRC model effectively predicts apCR in breast cancer patients after NAC. It can serve as an auxiliary tool for clinicians in identifying potential candidates for axillary de-escalation, offering support for personalized precision treatment. We present this article in accordance with the TRIPOD-AI reporting checklist.