COMPARATIVE ANALYSIS OF U-NET, U-NET++, TRANSUNET AND SWIN-UNET FOR LUNG X-RAY SEGMENTATION

Abstract

Medical image segmentation is a widely used task in medical image processing. It allows us to receive the location and size of the required instance. Several critical factors should be considered. First, the model should provide an accurate prediction of the mask. Second, the model should not require a lot of computational resources. Finally, the distribution between the false positive and false negative predictions should be considered. We provide the comparative analysis between four deep learning models, base U-Net and its extension U-Net++, TranUNet, and Swin-UNet for lung X-ray segmentation based on trainable parameters, DICE, IoU, Hausdorff Distance, Precision and Recall. CNN models with the smallest number of parameters show the highest DICE and IoU scores than their descendants on the limited-size dataset. Based on the experiment results provided in the article U-Nethas maximum DICE, IoU, and precision. It makes the model the most appropriate for medical image segmentation. SwinU-Net is the model with minimum Hausdorff Distance. U-Net++ has the maximum Recall.