Multi-scale Dual-Attention-Based U-Net for Breast Cancer Segmentation in Ultrasound Images

Ultrasound imaging is a significant and valuable assistant in reducing breast cancer mortality rate. It has high sensitivity, considered safe and cost-effective. This excellent performance caused the development of a sudden and growing interest in segmenting the breast ultrasound images to distinguish the abnormalities in them. However, this is a very challenging task due to the noisy nature of these images. This paper proposed a Multi-scale Dual-Attention based U-Net framework (MDAU-Net) for automatic breast ultrasound segmentation to address this issue. Our developed framework enhanced the quality of the extracted features by integrating different techniques such as multi-scale input representation, channel attention, spatial attention, blended attention, hybrid pooling and dilated convolutions. The experimental results and the comparison of the proposed model with the state-of-the-art segmentation methods on the well-known widely-used public ‘Dataset B’ demonstrate the effectiveness and competitiveness of the proposed algorithm.