Improved U-Net Based on Dual Attention Mechanism for Glottis Segmentation and Dysphagia Auxiliary Diagnosis

In today’s aging society, the proportion of elderly population is increasing year by year, and providing comprehensive care for the elderly has become an important issue. Among many aging diseases, dysphagia is a health threat that we often overlook, which if not detected and treated in time, can lead to aspiration pneumonia. Currently, the main detection method is usually through imaging of the throat and judgment by doctors. However, inexperienced doctors may make misjudgments. In order to avoid such situations, this study hopes to assist doctors in their diagnosis through effective image semantic segmentation technology. In the field of medical image semantic segmentation, the U-Net architecture has been proven to be a successful image segmentation architecture. The encoder-decoder technology in U-Net can effectively extract features and restore the original image. However, U-Net may lose important features during the downsampling process of feature extraction. Therefore, this study added a dual attention mechanism in the encoder, which effectively captures important features through position attention and channel attention in the image. In addition to the dual attention mechanism, this study added ResNet blocks in each encoder and decoder block to preserve feature information between downsampling and upsampling. Finally, this paper proves the effectiveness of these mechanisms through experiments and obtains good results.