Accurate segmentation of liver and tumors from medical imaging is critical for effective diagnosis and treatment planning in hepatic diseases. This paper introduces robust 3D deep learning architectures tailored for precise delineation of liver and tumor regions from Computed Tomography (CT) scans. Our approach is based on the encoder-decoder framework integrating specialized attention mechanisms and a Modified 3D Feature Pooling Module (MFPM) in U-Net architecture to handle the intrinsic challenges posed by the variable size & shape of tumors and the heterogeneous texture of liver tissues. Our methodology begins with a meticulous preprocessing stage that enhances image contrast and normalizes intensity distributions, followed by data augmentation to ensure robust model training. Leveraging the 3D spatial context, the proposed architecture employs attention-enhanced U-Net variants and a MFPM module, allowing for high-resolution feature extraction and multi-scale analysis essential for distinguishing complex anatomical structures. Experimental results obtained by investigation on Liver Tumor Segmentation Challenge (LiTS) dataset demonstrate that our models achieve the superior performance in terms of Dice Score Coefficient (DSC), Precision, Recall, and F2 with values 0.92, 0.91, 0.91, and 0.91, respectively for liver segmentation while the same for tumor lesions are 0.70, 0.68, 0.73, and 0.72, respectively. The proposed architectures exhibit exceptional capability in the segmentation of liver and tumors when compared with other contemporary models in terms of accuracy and efficiency. This work not only advances the state-of-the-art in medical image segmentation, but also suggests a scalable approach for other complex segmentation tasks in medical imaging.