Liver tumor segmentation is crucial in accurate diagnosis and effective treatment planning. The research investigates the performance of three known deep learning models: U-Net, AH-Net, and V-Net in liver tumor segmentation tasks, utilizing the MONAI framework for training and testing on a dataset of liver CT scans. The evaluation metrics include the Dice Similarity Coefficient (DSC), Intersection over Union (IoU) score, sensitivity, and specificity, commonly used in medical image segmentation. Our findings show distinct strengths and limitations across the models. U-Net, with its encoder-decoder architecture, demonstrated strong adaptability, achieving a Dice score of 0.92 and IoU score of 0.86, though it struggled to capture fine tumor details. V-Net, utilizing 3D convolutions and skip connections, excelled in volumetric segmentation, achieving the highest Dice score of 0.93 and IoU score of 0.87. However, AH-Net, despite its attention mechanism, showed inconsistencies, with a lower Dice score of 0.77 and IoU score of 0.69, indicating limitations in its effectiveness for this specific task.