摘要
随着医疗需求的持续增长,深度学习技术在医学图像自动分割领域展现出巨大的潜力。空间数据智能的发展为医学图像的精确分割提供了新的解决思路。U-Net作为医学图像分割领域最具影响力的网络架构,自2015年提出以来在各类医学影像任务中得到了广泛应用,其独特的编码器-解码器结构设计不仅为后续研究奠定了基础范式,更催生了大量改进网络。系统梳理了U-Net架构的重要发展里程碑:ResUNet通过残差连接解决了深层网络训练困难的问题,Attention-UNet引入自适应注意力机制提升了在跳跃连接中的特征选择精确度,而TransUNet和Swin-UNet则代表了将现代Transformer引入医学图像分割的2个关键阶段,展现了卷积神经网络(Convolutional Neural Network,CNN)与Transformer融合的巨大潜力。通过分析这些代表性网络的架构创新和性能突破,揭示了医学图像分割技术从纯CNN架构向CNN-Transformer混合架构演进的发展趋势。此外,探讨了现有技术面临的挑战,对未来空间数据智能的发展方向提供了见解,为该领域的进一步研究提供了参考。
With the increasing demands in healthcare,deep learning technologies have shown tremendous potential in the field of automatic medical image segmentation.The advancement in spatial data intelligence provides novel solutions for achieving high-precision segmentation in medical imaging.As the most influential/prominent network architecture,U-Net has been widely applied across various medical imaging tasks since its introduction in 2015.Its distinctive encoder-decoder structure design not only establishes a fundamental paradigm for subsequent research but also has inspired numerous network improvements.A systematic review of key advancements in U-Net architecture development is provided.ResUNet addresses the training challenges in deep networks by incorporating residual connections,which enhances model stability and convergence.AttentionUnet improves feature selection accuracy within the skip connections by introducing adaptive attention mechanisms.TransUnet and Swin-Unet represent two crucial stages in incorporating modern Transformer architectures into medical image segmentation,demonstrating the significant potential of CNN-Transformer fusion for improving feature representation and segmentation accuracy.The evolutionary trend of medical image segmentation technology is highlighted from conventional CNN-based architectures to CNN-Transformer hybrid architectures.Additionally,the challenges faced by existing technologies are addressed and insights into future development/research directions are provided,providing a comprehensive reference for further research in this field.
作者
刘紫权
史旭阳
胡海
马远萍
朱哲维
李珂
LIU Ziquan;SHI Xuyang;HU Hai;MA Yuanping;ZHU Zhewei;LI Ke(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China;Tianfu Institute of Research and Innovation,Southwest University of Science and Technology,Chengdu 610229,China;Science and Technology Informatization Division,Chengdu Public Security Bureau,Chengdu 610017,China)
出处
《无线电工程》
2024年第12期2765-2779,共15页
Radio Engineering
基金
四川省科技计划(2024NSFSC2040)
教育部高校产学研创新基金(2023IT257)
成都市技术创新研发项目(2024-YF05-01130-SN)
西南科技大学博士基金项目(23zx7136,23zx7135)。
