GSA-Net: Global Spatial Structure-Aware Attention Network for Liver Segmentation in MR Images With Respiratory Artifacts

Automatic liver segmentation is of great significance for computer-aided treatment and surgery of liver diseases. However, respiratory motion often affects the liver, leading to image artifacts in liver magnetic resonance imaging (MRI) and increasing segmentation difficulty. To overcome this issue, we propose a global spatial structure-aware attention model (GSA-Net), a robust segmentation network developed to overcome the difficulties caused by respiratory motion. The GSA-Net is an encoder-decoder architecture, which extracts spatial structure information from images and identifies different objects using the minimum spanning tree algorithm. The network's encoder extracts multi-scale image features with the help of an effective and lightweight channel attention module. The decoder then transforms these features bottom-up using tree filter modules. Combined with the boundary detection module, the segmentation performance can be further improved. We evaluate the effectiveness of our method on two liver MRI benchmarks: one with respiratory artifacts and the other without. Numerical evaluations on different benchmarks demonstrate that GSA-Net consistently outperforms previous state-of-the-art models in terms of segmentation precision on our respiratory artifact dataset, and also achieves notable results on high-quality datasets.