The precise segmentation of the liver and hepatic tumors from CT scans is a critical and complex step in transplantation and oncology workflows, as accurate volume measurements are essential for surgical planning, determining treatment strategies, and assessing therapeutic response. To overcome the limitations of time-consuming manual contouring which is prone to inter-observer variability and the inconsistent performance of existing automated tools in complex anatomical scenarios, we propose a novel deep learning framework designed for both high accuracy and clinical practicality. Our model integrates a modified EfficientNet-B0 backbone, enabling computationally efficient, multi-scale feature extraction while preserving fine spatial details essential for boundary delineation. It is further enhanced with a Hybrid Attention (HA) mechanism that dynamically prioritizes salient liver and tumor regions, alongside dedicated Boundary Refinement (BR) modules that iteratively sharpen critical margins in challenging areas of low contrast or irregular morphology. Quantitative evaluation on the publicly available, expert-annotated 3D-IRCADb dataset demonstrates that our method achieves superior segmentation accuracy, with Dice similarity coefficients of 96.37% for liver and 94.27% for liver tumors, while also excelling in boundary-specific metrics compared to current state-of-the-art approaches. These results validate the model’s robustness across diverse clinical presentations and highlight its strong potential for integration as a reliable decision-support tool, ultimately aiming to improve the efficiency, reproducibility, and precision of liver imaging analysis in routine clinical practice.