DA-Net: Deep attention network for biomedical image segmentation

Deep learning-based image segmentation techniques are of great significance to biomedical image analysis and clinical disease diagnosis, among which U-Net is one of the classic biomedical image segmentation algorithms and is widely used in the field of biomedicine. In this paper, we propose an improved triplet attention module and embed it into the U-Net framework to form a novel deep attention network, called DA-Net, for biomedical image segmentation. Specifically, an additional layer is stacked into the original U-Net, resulting in a six-layer U-shaped network. Then, the double convolution module of the U-Net is replaced with a composite block which consists of the improved triplet attention module and the residual concatenate block, to obtain abundant valuable features effectively. We redesign the network structure to increase its width and depth and train our model with the pixel position aware loss, realizing the synchronous increase of the mean IoU value and average Dice index. Extensive experiments have been carried out on two publicly available biomedical datasets, including the 2018 Data Science Bowl (DSB) and the international skin imaging collaboration (ISIC) 2018 Challenge, and a self-built fetal cerebellar ultrasound dataset from Affiliated Hospital of Jiangsu University, named JSUAHCerebellum. The mIoU and mDice of DA-Net can reach 87.45 % and 92.98 % on the JSUAHCerebellum, 87.36 % and 91.37 % on the 2018 Data Science Bowl, and 86.75 % and 91.34 % on the ISIC-2018 Challenge, respectively. Experimental results demonstrate that our DA-Net achieves promising performance in terms of segmentation robustness and generalization ability.