UTSN-net: medical image semantic segmentation model based on skip non-local attention module

Abstract

The semantic segmentation task of medical image is to segment the focus, organ or substructure of human body in medical image. It plays an important role in locating and identifying the diseased area and making medical plan. In various medical image segmentation tasks, the U-shaped architecture has achieved great success. Transunet introduces Transformer with global attention mechanism into the U-shaped architecture, which overcomes the inherent limitations of convolution, but because it still continues the original skip connections structure, it will bring the strong noise from features in the shallow network into the high semantic features of the deep network, thus affecting the segmentation accuracy. UTSN-net model based on the combination of Transformer and nonlocal attention mechanism is proposed. UTSN-net uses Transformer to overcome the inherent limitations of convolution, and introduces the skip connections module based on nonlocal attention mechanism into the U-shaped network, which can comprehensively consider the deep features with global context information and the shallow features with accurate high-resolution positioning information to improve the accuracy of segmentation results. Experiments on synapse multi-organ abdominal CT dataset verify that UTSN-net has better semantic segmentation performance.