Predicting axillary lymph node metastasis in breast cancer patients using CNN-GCN on DCE-MRI: a multicenter study

BACKGROUND: To develop a deep learning model combining a convolution neural network (CNN) and a graph convolution network (GCN) based on dynamic contrast-enhanced (DCE) MRI for predicting axillary lymph node (ALN) metastasis in breast cancer patients while also aiming to explore the underlying biological mechanism by using RNA sequencing (RNA-seq) data. METHODS: We retrospectively collected DCE-MRI and clinical data from 1002 patients across four centers and one public dataset. Various CNN-GCN models were trained on tumor and ALN images and compared to radiomics models, the MSKCC model, and radiologists. RNA-seq data from 11 patients were used to explore associated biological pathways. Model performance was evaluated by accuracy, sensitivity, specificity, AUC, and DeLong test. RESULTS: Participants were divided into a training set (n = 742, mean age: 53 years ±10 [SD]), an internal test set (n = 83, 53 years ±10), external test set 1 (n = 110, 50 years ±9) and external test set 2 (n = 67, 54 years ±11). The optimal CNN-GCN model, HRNet-GCN_tumor+ALN, achieved an AUC of 0.873 in the internal test set outperforming the LR_tumor+ALN (AUC: 0.790) and MSKCC models (AUC: 0.726) (DeLong test, p < 0.05). Radiologists’ performance improved with HRNet-GCN_tumor+ALN (in both the external test set 1 and 2, p < 0.05). High-risk group associated with pathways such as ribosome, synapse organization, and muscle contraction. CONCLUSIONS: The proposed CNN-GCN fusion deep learning model showed good performance for preoperatively predicting ALN status in breast cancer patients.