Residual-fusion Network for Multi-modality Breast Cancer Image Analysis

Breast ultrasound (BUS) imaging is typically used for breast cancer detection techniques for reasons such as nonionizing, cost-effective, and noninvasive. This ultrasound image has several modalities. The most representative modality is B-mode, which is a modality that visualizes the overall morphologic information. This information is important for detecting the lesion area. Strain elastography (SE) mode is the modality that shows tissue stiffness. Therefore, we use B-mode for identifying the lesion and supplement information about tissue stiffness through SE-mode to more accurately predict the area and border of the lesion. The segmentation model followed the existing U-Net structure with the fusion modules. With these modules, information about the two modalities can be mixed at each layer. When segmentation is completed, the input image is cropped based on the predicted area, and classification is performed through an encoder of the same structure. With this proposed method, Dice score for the segmentation task is 78.23%, and the accuracy for the classification task is 98.46%.