Medical image segmentation is an essential field of image analysis that processes and extracts information using state-of-the-art deep learning techniques. However, there are various challenges to overcome. One of these challenges is the class imbalance for medical image datasets in which lesions often occupy a much smaller volume than the background. Thus, deep learning algorithms vary in robustness to class imbalance in medical images. Moreover, most training for standard medical datasets uses loss functions for segmentation based on cross-entropy loss, dice loss, or a combination of both. Selecting an optimal loss function affects the performance of the segmentation results. To address these topics, this research has proposed integrating focal loss into a hierarchical framework to improve these traditional loss functions. The proposed method is evaluated on a medical imaging dataset related to the abdominal cavity, known for its imbalances. A comparative analysis is conducted between the original LeViT-UNet model and its modified version using the new model. Results show that the modified model significantly outperforms the original one. It indicates the potential of focal loss integration as an effective solution for improving segmentation performance in medical imaging.