The Evolution of U-Net Architectures in Medical Image Segmentation

School of Computer Science and Technology, Donghua University, Shanghai, 201620, China

^* Corresponding author: 221310229@mail.dhu.edu.cn

Abstract

As an important field of deep learning, image segmentation has developed rapidly in recent years. Medical image segmentation has always been an important application scenario of image segmentation, and many models with excellent performance have emerged. U-Net has become the focus of research and application recently because of its easy to understand structure and excellent performance. This paper focuses on the structure of U-Net itself and explains the reasons for U-Net's excellent performance as well as some defects. At the same time, this paper also introduces the improvements and adjustments made by different researchers to solve the problems encountered in the practical application of U-Net, including structural improvements, such as the adjustment of modules and the replacement of convolution methods, and non-structural improvements, such as the optimization of data sets and the improvement of loss functions. Finally, the prospects and suggestions for the future development and application of U-Net were put forward.