SEG-LUS: A novel ultrasound segmentation method for liver and its accessory structures based on muti-head self-attention

Lei Zhang, Xiuming Wu, Jiansong Zhang, Zhonghua Liu, Yuling Fan, Lan Zheng, Peizhong Liu, Haisheng Song, Guorong Lyu

Abstract: Highlights•We propose SEG-LUS, a semantic segmentation model for liver ultrasound diagnostic and therapeutic processes. It is the sole model applicable to LUSS recognition, effectively representing key anatomical structures during clinical scanning.•Introducing a hybrid attention mechanism, Cross Shift-window MSA (CSW-MSA), combined with UUF for liver ultrasound analysis, achieves top performance on an LUSS dataset with eight key anatomical structures.•Compared with seven other leading segmentation methods, we achieve a 4.5% lead over the baseline average. Results include mPA 85.05%, mDice 82.60%, mIOU 74.92%, and mASD 0.31. This provides an effective design reference for automated computer-aided modeling based on liver ultrasound data.