U-Net vs. TransUNet: Performance Comparison in Medical Image Segmentation

Abstract

Image segmentation is a fundamental task in computer-aided diagnosis systems. Correct image segmentation can help healthcare professionals to have better arguments to define a diagnosis and a possible treatment. Among the existing methods and techniques for this task is the TransUNet architecture. This architecture comes from the U-Net architecture, its predecessor, and has the characteristic of incorporating transformers in its encoder. This paper analyzes TransUNet and contrasts its performance with its predecessor. For this purpose, the architectures were implemented, trained, and tested using the Medical Segmentation Decathlon dataset, which provides new challenges and tasks different from those usually addressed. The experimental results reveal the superiority of TransUNet over U-Net in both spleen and left atrium segmentation by reducing the loss in 83% and 72% in the training and testing set, respectively, in the spleen task; and 94% and 92% in the left atrium task. Likewise, in the dice score, TransUNet achieved a superiority of 3.42% and 0.34% for the spleen and left atrium task, respectively, over U-Net.