A novel U-shaped AFS-Net with attention fusion module for automated liver and tumor segmentation in abdominal CT images

Accurate segmentation of the liver and liver tumors on abdominal computed tomography is essential for diagnosis, staging, and treatment planning in hepatocellular carcinoma. However, manual delineation is time-consuming, labor-intensive, and subject to observer variability. Existing deep learning methods still face challenges related to tumor heterogeneity, indistinct tumor-liver boundaries, and low-contrast lesions. We developed AFS-Net, a U-shaped deep learning framework for automated liver and tumor segmentation. The model incorporates an attention fusion module in skip connections to adaptively integrate encoder-decoder features and a multi-scale deep supervision strategy to improve training. AFS-Net was evaluated on the public Liver Tumor Segmentation dataset and compared with 3D U-Net, ResUNet, and nnUNet. Performance was assessed using the Dice coefficient, the 95th percentile Hausdorff distance, and the average symmetric surface distance. Qualitative evaluation was also performed by two radiologists on the cases. AFS-Net achieved the best overall performance among all compared models. For liver segmentation, it achieved a Dice coefficient of 96.5%, a 95th percentile Hausdorff distance of 18.9 mm, and an average symmetric surface distance of 2.1 mm. For tumor segmentation, it achieved a Dice coefficient of 78.2%, a 95th percentile Hausdorff distance of 18.5 mm, and an average symmetric surface distance of 3.2 mm. Compared with nnUNet, AFS-Net improved tumor Dice and reduced boundary-based errors. In radiologist-based evaluation, accepted contours showed consistently better quantitative performance than rejected contours. AFS-Net is an effective method for automated liver and tumor segmentation on computed tomography images. It improves both overlap accuracy and boundary delineation, particularly for tumor segmentation, and may support computer-assisted diagnosis and treatment planning.