Publication

Abstract : Liver cancer is a prevalent malignant tumor with high global incidence and mortality rates. It is still very difficult to accurately segment the liver and related tumors in contrastenhanced CT (CECT) imaging. The major class imbalance between the two regions, the high level of morphological diversity in tumors, and the inability to distinguish lesions from surrounding tissues by contrast are the main causes of this difficulty. Although U-Net architectures are commonly used for these tasks, the precise effects of different encoder backbones on performance across various datasets have not yet been thoroughly investigated. In this work, we present a methodical assessment of different encoder networks for multi-class segmentation using two different datasets: the public LiTS dataset and the Primary Liver Cancer CECT Imaging Dataset. In comparison to a typical vanilla U-Net baseline, we evaluated U-Net models integrated with ResNet-50, ResNet-101, EfficientNet-B0, and EfficientNet-B4. All architectures were optimized using a triple-hybrid loss function that included Tversky, Dice, and Cross-Entropy components in order to reduce the severe volumetric difference between the localized tumor regions and the expansive liver. Strict patientlevel data partitioning was used to guarantee equitable generalizability throughout the assessment process. The Hausdorff Distance, Average Surface Distance (ASD), and Dice Similarity Coefficient (DSC) were used to gauge the model's effectiveness. According to our research, EfficientNet-based encoders routinely perform better than alternative backbones, with the EfficientNetB4 variant being the best model. The mean Dice score for this configuration was 86.27 % on the CECT dataset and 86.83 % on the LiTS dataset.