关键词: 医学图像分割, U-Net, 结构改进, Transformer, 深度神经网络, 技术综述

Abstract: The accurate segmentation of lesions in medical images plays a key role in the physician's search for the cause of disease and the formulation of treatment plans. The development of computer vision technology has led to the derivation of various model architectures for deep learning in medical image segmentation. The U-Net architecture is widely used for its clever skip connection, easy-to-optimize module design has become a benchmark model in this field. However, U-Net network, with Convolutional Neural Network (CNN) as its backbone, is only good at acquiring local features in terms of long-term modeling dependencies, and various CNN-based methods lack the interpretation of long-term image correlations in performing segmentation tasks to extract global features. In order to help scholars in this field to understand the development history and research status of U-Net, in the text provides a problem-oriented overview of the U-Net improvement work in the last seven years. First, the U-Net and its various improved models are narrated from the perspective of improving the structural position, and the research purpose, innovative design and shortcomings of each work are discussed. Secondly, the combination method of Transformer and U-shaped network is analyzed to obtain research trends for improvement work. Finally, comparative experiments are conducted on Synapse and ACDC datasets, and the experimental analysis and visualization results show that the Transformer method has a significant advantage in segmentation accuracy, especially the way of combining hybrid network sub-blocks, which ensures the model performance while taking into account the efficiency, and proves that this class has a broad development prospect and research value.

Key words: medical image segmentation, U-Net, structural improvement, transformer, deep neural network, technology overview