Multichannel deep learning-based MRI model for predicting breast cancer axillary lymph node invasion: a comparative
Lingsong Meng1,2,3, Yuxia Zhang1, Xin Zhao3
1Department of Medical Technology, Shangqiu Medical College, Shangqiu, China.
Background:
The accurate evaluation of axillary lymph node (ALN) status before surgery is critical to formulate appropriate surgical strategies and determine whether axillary lymph node dissection (ALND) should be performed in breast cancer (BC) patients. This study aimed to develop a multichannel multiscale deep learning (DL) model for the preoperative prediction of ALN metastasis for BC patients using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and to compare its performance with that of the Node Reporting and Data System (Node-RADS).
Methods:
A total of 735 patients were included in this two-center retrospective study. A DL framework integrating features from axial, sagittal, coronal, and multiplanar reformat DCE-MRI images was constructed. ResNet101 served as the feature extractor, followed by a transformer-based fusion module. Model interpretability was enhanced with gradient-weighted class activation mapping (Grad-CAM) and SHapley additive exPlanations (SHAP). Diagnostic performance was evaluated using the area under the curve (AUC), sensitivity, and specificity, and compared with the Node-RADS in an external validation cohort (n=148).
Results:
The proposed model achieved promising results with AUC values of 0.959 [95% confidence interval (CI): 0.942-0.976] in the training cohort, 0.885 (95% CI: 0.839-0.931) in the internal validation cohort, and 0.908 (95% CI: 0.862-0.954) in the external validation cohort. The predictive ability of the model remained stable across diverse patient subgroups stratified by age, tumor size, and Breast Imaging Reporting and Data System (BI-RADS) category. In the external validation cohort, the model demonstrated diagnostic accuracy comparable to that of the Node-RADS (AUC: 0.908 vs. 0.909, P=0.971), with no significant differences in sensitivity, specificity, or accuracy (all P>0.05). Agreement between the two approaches was moderate (kappa =0.523).
Conclusions:
The developed DL model provides accurate and robust preoperative prediction of ALN metastasis and its performance was comparable to that of the Node-RADS. It could serve as a clinical decision-support tool to reduce unnecessary ALN dissections.