Deep Residual U-Net Framework for Liver Tumor Segmentation on CT Images

Accurate segmentation of liver and tumors from computed tomography images is essential for reliable diagnosis, treatment planning, and surgical guidance in hepatic diseases. However, challenges such as low contrast, irregular tumor morphology, and intra-patient variability often hinder the performance of conventional segmentation algorithms. In this study, we propose a deep-learning-based hybrid model that integrates U-Net with ResNet backbones such as ResNet18, ResNet34, and ResNet50 to enhance segmentation accuracy and efficiency. The models were trained and evaluated on two publicly available datasets of LiTS17 and 3DIRCADb using 2D and 2.5D axial slices extracted from 3D CT volumes. Experimental results show that the U-Net + ResNet34 configuration consistently achieved superior performance obtaining a Dice coefficient of 0.9708 and an IoU of 0.9435 on the LiTS17 dataset, and a Dice Coefficient of 0.9284 and an IoU of 0.8706 across the 3DIRCADb dataset. The model maintained a lower training time compared to deeper networks. The marginal performance improvements are observed across models, and statistical significance test such as the Wilcoxon signed-rank test with confidence intervals shows that gains are consistently significant, which indicates the importance of cautious interpretation. The findings underline the value of incorporating residual learning into an encoder–decoder model for precise medical image segmentation. The proposed approach shows strong potential for clinical adoption and provides a scalable foundation for future research in 3D volumetric or hybrid 2.5D and multimodal image segmentation.