Abstract
Liver tumor segmentation from CT images remains challenging due to large variations in lesion scale, blurred boundaries, low tissue contrast, and the high computational cost of existing deep learning models. This study aims to develop a lightweight yet accurate segmentation network suitable for clinical deployment. We propose a Multi-scale Multi-attention Boundary-aware Network (M²B-Net) based on a U-shaped encoder-decoder architecture, integrating four modules: Multi-Dimensional Spatial-Location Attention (MDSLA) for feature enhancement across encoder scales, Manhattan Self-Attention (MaSA) for global dependency modeling, Multi-Scale Feature Refinement Module (MSFRM) for cross-scale feature alignment, and Boundary-Convolution Attention Module (BCAM) for edge detail enhancement. A weighted composite loss function combining cross-entropy and Dice loss is used. Experiments were conducted on the public LiTS dataset. M²B-Net achieved a Dice coefficient of 0.77 ± 0.13, volumetric overlap error of 0.35 ± 0.15, average symmetric surface distance of 2.99 ± 2.03 mm, and maximum symmetric surface distance of 5.20 ± 3.07 mm. The model contains 21.7 million parameters and 28.5 GFLOPs, with a training time of 4.1 h and testing time of 37 s per case. Ablation and comparative experiments confirmed the contribution of each module and showed superior performance over SegNet, TD-Net, SBC-Net, and RIS-UNet. M²B-Net effectively addresses multi-scale adaptation, weak boundary capture, and global context modeling in liver tumor segmentation while maintaining a lightweight architecture, demonstrating strong potential for rapid and accurate clinical deployment in resource-constrained settings.