Multi-Attention Enhanced Segmentation Network for Medical Image

In clinical medicine, accurate segmentation of organ contour is the prerequisite for assisting doctors in diagnosis and precise radiotherapy. In order to improve the problem that U-Net network caused by the local nature of convolution operation can only acquire local features in medical image segmentation, and it is difficult to capture global context information, an multi-attention enhanced segmentation network (MAESN et) for medical image. The network is based on a U-Net network and consists of a convolutional block attention module (CBAM) and an attention-enhancing skip connection (AESC) module. The CBAM module recalibrates and enhances the features from two dimensions of channel and spatial perform adaptive feature optimization, the attention enhancement skip connection module replaces the skip connection, fuses the multi-level features from the encoder, effectively integrates the feature information of different scales, bridges the semantic gap between the encoder and the decoder, and strengthens the global dependency. Comparison experiments and ablation experiments were performed on the own dataset, and the results showed that MAESNet achieved 81.60% in DSC and 26.04 in HD, respectively, effectively improving the segmentation accuracy.