State Classification of Axillary Lymph Nodes via Dual-Modal Adaptive Fusion

Breast cancer, being the most prevalent malignancy among women globally, presents a significant public health challenge due to its high incidence and mortality rates. To aid clinicians in a more precise preoperative assessment of the axillary status in early-stage breast cancer patients, a framework for classifying axillary lymph node status was designed based on the transformer. Firstly, model parameters were adjusted to determine the optimal fusion period for B-mode ultrasound and shear wave elastography dual-modal image features. To fully reveal the intrinsic medical clues contained in the dual-modal images, two feature fusion schemes, namely, mid-term adaptive interaction fusion and terminal adaptive fusion, were proposed. To balance the contribution of dual-modal images to axillary lymph node metastasis (ALNM) prediction, an adaptive fusion binary cross-entropy loss function and Focal loss function were designed as a composite loss function. Experimental results demonstrated that the proposed network performed excellently in the task of predicting early breast cancer patients, achieving a 92% accuracy and 0.94 AUC, surpassing existing deep learning classification models for breast cancer.