Abstract: With the growing demand for high-precision diagnostics in society, automated medical image segmentation technologies have assumed a pivotal role in modern medical practice. Although convolutional neural network (CNN) has demonstrated excellent performance in medical image segmentation, its inherent limitations have prompted many researchers to incorporate Transformers into this domain to address CNN’s shortcomings in global contextual learning. This paper first reviews the structure of Transformers and their variants, analyzing their integration and application in medical image segmentation tasks. Focusing on five major segmentation tasks: cardiac, brain, lung, abdominal and other regions, it summarizes the research progress combining U-Net and other models, highlighting their advantages in capturing multi-scale features, improving segmentation accuracy, and addressing the complexity of diverse anatomical structures. Additionally, existing studies are discussed, emphasizing the need for further in-depth research to advance the development of medical image segmentation technologies.

Key words: deep learning, Transformer, medical image segmentation, convolutional neural network (CNN)

摘要： 随着社会对高精度诊断的需求持续攀升，自动化医学图像分割技术于现代医疗实践中占据着关键地位，尽管卷积神经网络（convolutional neural network，CNN）在医学图像分割方面表现优异，但由于其存在一定局限性，许多学者遂将Transformer引入医学图像分割领域，以弥补CNN在全局上下文学习层面的欠缺。综述了Transformer及其变体结构，并分析了它们在医学图像分割任务中的结合应用。从心脏、大脑、肺部、腹部和其他部位这五个主要分割任务范畴，归纳了基于U-Net以及其他模型相结合的研究进展，指出其在捕捉多尺度特征、提升分割精度以及应对不同解剖结构的复杂性方面所具备的优势。对现有研究工作进行了讨论，未来需再持续深入研究，以促进医学图像分割技术的发展。

关键词: 深度学习, Transformer, 医学图像分割, 卷积神经网络（CNN）