Liver Segmentation Using Modified CAG-SwinUNet with Explainability

Accurate liver segmentation from computed tomography (CT) images is crucial for clinical applications such as tumor detection and surgical planning but remains challenging due to anatomical complexity and imaging variability. Existing deep learning models, struggle with ambiguous liver boundaries, noise sensitivity, and weak feature integration across scales, leading to segmentation errors. This study introduces Cross-Attention Gate-Shifted Window U-Net (CAG-SwinUnet), an enhanced Swin-UNet variant that incorporates a Cross-Attention Gate (CAG) in skip connections to selectively refine feature fusion. Unlike traditional concatenation, CAG dynamically enhances encoder features based on decoder context, integrating residual connections and output projection to balance local and global information. Extensive evaluation on Liver Tumor Segmentation (LiTS) and Segmentation of the Liver Competition 2007 (SLIVER07) demonstrates state-of-the-art performance, achieving 97.75% Dice Similarity Coefficient (DSC) and 2.40 mm Hausdorff Distance (HD) on LiTS, and 96.65% DSC and 3.10 mm HD on SLIVER07, respectively. To enhance explainability, gradient-weighted class activation mapping, provide visual insights into the model’s decision-making process, ensuring transparency and reliability in liver segmentation.