Deep Learning-Based Liver Tumor Segmentation: A Comparative Study of U-Net Variants for Medical Imaging Analysis

Liver tumor segmentation(LiTS) is a critical task in medical imaging analysis, playing a key role in the early detection, diagnosis, and treatment planning of liver diseases. Accurate segmentation of liver tumors from medical images is vital for quantifying tumor size, monitoring disease progression, and assessing treatment outcomes. However, manual segmentation is time-consuming and subjective, highlighting the need for automated and reliable deep learning-based methods. This study aims to evaluate and compare the performance of deep learning architectures: U-Net and its variants (U-Net, U-Net++, and newly introduced U-Net Square, and U-Net Cube) for segmenting liver tumors using medical imaging data. Each variant’s unique architectural modifications are examined, emphasizing their potential benefits for liver tumor segmentation tasks. The efficacy and performance of these models for detecting liver tumors is analyzed using metrics like Dice Coefficient. It is observed that basic U-Net outperforms other architectures in liver tumor segmentation. This study further provides potential implications for enhanced diagnostic accuracy and treatment planning in clinical practice. The results contribute valuable insights into the strengths and limitations of U-Net variants, assisting researchers in selecting optimal deep-learning models for liver tumor segmentation.