Deeply supervised breast cancer segmentation combined with multi-scale and attention-residuals

Given the low accuracy of delineation of the infiltration area of breast cancer lesions in DCE-MRI, variable structure and shape, large intensity heterogeneity changes, and low boundary contrast, the automatic segmentation of breast cancer lesions has the problems of low accuracy and mis-segmentation. For this reason, a two-stage breast cancer image segmentation framework is constructed, and a breast cancer lesion segmentation model UTB-net is proposed to integrate multi-scale and non-local at the encoding path and the end, respectively, which constructs attention-residuals in the decoding module. First, the benchmark U-net network model is used to achieve a rough delineation of the breast area, eliminating the influence of unrelated tissues, such as chest muscle, fat, and the heart, on breast tumor segmentation in the image. Then, based on the extracted ROI results, a multi-scale information fused and non-local module is constructed in the coding path of the model. Finally, an attention-residual hybrid decoding module is constructed in the decoding path, and a deep supervision mechanism is introduced to improve the segmentation accuracy of breast tumor lesions. Experimental results show that breast tumor segmentation indexes DICE, IOU, SEN, and PPV increase by 4%, 4.78%, 5.92%, and 3.94% respectively, in comparison with the U-Net benchmark model. The proposed model not only improves the segmentation results of breast cancer but also reduces the small-area mis-segmentation and calcification segmentation. © 2021, Science Press. All right reserved.