Synergistic attention U-Net for sublingual vein segmentation

The tongue is one of the most sensitive organs of the human body. The changes in the tongue indicate the changes of the human state. One of the features of the tongue, which can be used to inspect the blood circulation of human, is the shape information of the sublingual vein. Therefore, this paper aims to segment the sublingual vein from the RGB images of the tongue. In traditional segmentation network training based on deep learning, the resolution of the input image is generally resized to save training costs. However, the size of the sublingual vein is much smaller than the size of the tongue relative to the entire image. The resized inputs are likely to cause the network to fail to capture target information for the smaller segmentation and produce an “all black” output. This study first pointed out that the training of the segmentation of the sublingual vein compared to the tongue segmentation is much more difficult through a small dataset. At the same time, we also compared the effects of different input sizes on small sublingual segmentation. In response to the problems that arise, we propose a synergistic attention network. By dismembering the entire encoder–decoder framework and updating the parameters synergistically, the proposed network can not only improve the convergence speed of training process, but also avoid the problem of falling into the optimal local solution and maintains the stability of training without increasing the training cost and additional regional auxiliary labels.