Comparative Evaluation of U-Net-Based Models for Lung Lesion Segmentation in CT Images

Lung cancer remains one of the leading causes of cancer-related deaths worldwide, and early detection plays a crucial role in improving treatment success and patient survival. In this study, the problem of lesion segmentation in lung computed tomography (CT) images was addressed, and the performance of different U-Net-based deep learning architectures was comparatively evaluated. The dataset used in this study consisted of 22,782 CT slices obtained from 223 patients. Among these images, 3,584 slices contained lesions, while 19,198 slices were healthy images. To investigate the impact of class imbalance in the dataset on model performance, three different experimental scenarios were applied: using only lesion-containing images, using all images together, and using a balanced training dataset. Segmentation performance was analyzed using both image-based and patient-based evaluation approaches. The results indicate that data distribution has a significant impact on segmentation performance. In particular, balancing the training dataset improved the results, especially in patient-based evaluations.