Optimized U ³ -Net framework for multi-class liver and tumor segmentation: a comparative study with boundary-aware clinical metrics

Abstract

Accurate segmentation of liver and tumor regions in Computed Tomography (CT) scans is fundamental for the effective diagnosis and surgical planning of hepatic malignancies. This study evaluates and compares three sophisticated Convolutional Neural Network (CNN) architectures—U-Net, U 2 -Net, and U 3 -Net—for the automated multi-class segmentation of background, liver, and tumor tissues. To address the inherent challenge of class imbalance and the scarcity of lesion samples, we implemented a strategic oversampling technique combined with extensive data augmentation. The models were optimized using a hybrid loss function (integrating Focal and Dice loss) to enhance sensitivity toward small and irregular tumor boundaries. Quantitative assessment was performed using a comprehensive suite of metrics, including overlap-based measures such as Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) as well as boundary-based clinical measures including the 95th percentile Hausdorff Distance (HD95), Average Symmetric Surface Distance (ASSD), and Relative Absolute Volume Difference (RAVD). Experimental results on the LiTS dataset demonstrate that U 3 -Net achieves superior performance, resulting in a Dice score of 0.97 for Liver Segmentation (LS) and 0.95 for Tumor Segmentation (TS).

Ähnliche Arbeiten