Accurate Segmentation of Imbalanced Medical Images: A Comparative Study

Abstract

Medical image segmentation is a fundamental task for accurate automated diagnosis, treatment planning, and clinical decision-making. This study presents a comparative evaluation of the LeViT-UNet model a convolutional encoder decoder network enhanced with transformer blocks on imbalanced computed tomography (CT) datasets. Two loss functions were investigated: the traditional Focal Loss and the composite Focal-Tversky Loss. The model was trained and validated on annotated CT slices exhibiting high class imbalance to assess segmentation accuracy and convergence stability. Experimental results reveal that training with Focal Loss enables faster convergence and achieves higher Dice and Jaccard scores during early epochs by emphasizing challenging samples. In contrast, the Focal-Tversky Loss achieves a better trade-off between sensitivity and specificity, leading to improved stability and generalization across imbalanced data. These findings underscore the importance of selecting task-specific loss functions for medical image segmentation and demonstrate that integrating LeViT-UNet with Focal-Tversky Loss provides a robust and consistent framework suitable for clinical applications demanding high precision.