Early identification of axillary lymph node metastasis in patients with primary breast cancer based on 3D deep learning radiomics nomogram

Background Axillary lymph node metastasis of breast cancer is vital for early treatment decision-making, but puncture biopsy may induce a series of adverse reactions in patients. This study aimed to noninvasively identify axillary lymph node metastasis (ALNM) using a deep learning radiomics nomogram model based on Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI). Methods Clinical and MRI data of 396 patients with invasive breast cancer (2018–2024) were retrospectively analyzed and split into training (n = 277) and test (n = 119) sets. Radiomics and ResNet50 deep learning features were extracted and selected via LASSO regression model. Eight models were constructed and interpreted by SHAP and Grad-CAM. Multivariate regression was performed to identify independent risk factors for ALNM. A (Deep Learning Radiomics) DLR-score-based nomogram was developed and assessed by area under the curve (AUC), sensitivity, specificity and accuracy. Results The clinical model yielded AUCs of 0.704 and 0.642, the radiomics model 0.813 and 0.804, and the DLR model 0.903 and 0.816 in the training and test sets, respectively. Carcinoma in situ, tumor location and pathological type were independent risk factors for axillary lymph node metastasis. The integrated nomogram achieved an AUC of 0.921, sensitivity of 0.821, specificity of 0.870, and accuracy of 0.848 in the training set, and 0.834, 0.821, 0.714, and 0.765 in the test set. Conclusion The nomogram based on DCE-MRI deep learning radiomics can effectively predict ALNM in breast cancer patients.