《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (12): 4037-4044.DOI: 10.11772/j.issn.1001-9081.2024111673

• 多媒体计算与计算机仿真 • 上一篇    下一篇

改进TransUNet的高效通道注意力医学图像分割网络

邓酩1,2, 徐锦凡2, 肖洪祥2, 谢晓兰2   

  1. 1.广西嵌入式技术与智能系统重点实验室(桂林理工大学),广西 桂林 541006
    2.桂林理工大学 计算机科学与工程学院,广西 桂林 541006
  • 收稿日期:2024-11-27 修回日期:2025-04-11 接受日期:2025-04-16 发布日期:2025-04-18 出版日期:2025-12-10
  • 通讯作者: 肖洪祥
  • 作者简介:邓酩(1979—),男,湖南永州人,副教授,硕士,主要研究方向:图像处理、智能计算
    徐锦凡(1998—),男,江西上饶人,硕士研究生,主要研究方向:医学图像语义分割
    肖洪祥(1965—),男,湖北武汉人,教授,硕士,CCF会员,主要研究方向:智能检测
    谢晓兰(1974—),女,广西桂林人,教授,博士,主要研究方向:云计算、大数据。
  • 基金资助:
    国家自然科学基金资助项目(62262011);广西重点研发计划项目(桂科AB23049001)

Medical image segmentation network based on improved TransUNet with efficient channel attention

Ming DENG1,2, Jinfan XU2, Hongxiang XIAO2, Xiaolan XIE2   

  1. 1.Guangxi Key Laboratory of Embedded Technology and Intelligent Systems (Guilin University of Technology),Guilin Guangxi 541006,China
    2.College of Computer Science and Engineering,Guilin University of Technology,Guilin Guangxi 541006,China
  • Received:2024-11-27 Revised:2025-04-11 Accepted:2025-04-16 Online:2025-04-18 Published:2025-12-10
  • Contact: Hongxiang XIAO
  • About author:DENG Ming, born in 1979, M. S., associate professor. His research interests include image processing, intelligent computing.
    XU Jinfan, born in 1998, M. S. candidate. His research interests include semantic segmentation of medical images.
    XIAO Hongxiang, born in 1965, M. S., professor. His research interests include intelligent detection.
    XIE Xiaolan, born in 1974, Ph. D., professor. Her research interests include cloud computing, big data.
  • Supported by:
    National Natural Science Foundation of China(62262011);Guangxi Key Research and Development Program(GuiKe AB23049001)

摘要:

医学图像分割在计算机辅助诊断和手术导航等临床应用中起着至关重要的作用,旨在从复杂的医学影像中精准提取不同器官和病灶。然而,现有的U型网络结构在实际应用中存在跳跃连接信息冗余大和计算量高等问题。为了解决这些问题,提出一种轻量化医学图像分割网络ES-TransUNet (Efficient channel attention and Simple-TransUNet)。该网络在编码器中通过引入十字交叉注意力(CCA)机制捕捉图像中的长距离依赖关系,并优化Transformer中的多头注意力结构,从而使模型轻量化,在解码器中引入动态上采样(Dysample)模块提升上采样效率;同时为了减少跳跃连接中的信息冗余,引入简单上下文Transformer(SCOT)块对冗余特征进行过滤。在Synapse多器官分割和ACDC数据集上的实验结果表明,ES-TransUNet相比TransUNet分别取得了2.37和1.57个百分点的Dice相似系数(DSC)提升,并在Synapse数据集上使Hausdorff距离(HD)降低了约9.69。此外,所提网络与现有最先进的医学分割模型的对比结果表明,ES-TransUNet在保持较高分割精度的基础上,显著降低了模型的参数量和计算复杂度,并提高了推理效率。可见,该网络更满足实时医学图像分割的实际需求。

关键词: 医学图像分割, U-Net, 轻量化, Transformer, 跳跃连接, 注意力机制

Abstract:

Medical image segmentation plays a crucial role in clinical applications such as computer-aided diagnosis and surgical navigation, aiming to extract different organs and lesions from complex medical images accurately. However, the existing U-shaped network architecture suffers from the problems such as high information redundancy in skip connections and high computational complexity. To address these challenges, a lightweight medical image segmentation network named ES-TransUNet (Efficient channel attention and Simple-TransUNet) was proposed. In the network, the Criss-Cross Attention (CCA) mechanism was introduced in the encoder to capture long-range dependencies and the multi-head attention structure in Transformer was optimized, so as to lighten the model. Dynamic upsampling (Dysample) module was introduced in the decoder to improve upsampling efficiency. At the same time, in order to reduce the information redundancy in skip connections, the Simple COntextual Transformer (SCOT) block was introduced to filter out redundant features. Experimental results on the Synapse multi-organ segmentation and ACDC datasets demonstrate that ES-TransUNet achieves 2.37 and 1.57 percentage points improvements, respectively, in Dice Similarity Coefficient (DSC) compared to TransUNet; and reduces the Hausdorff Distance (HD) by 9.69 approximately on the Synapse dataset. Additionally, the results of comparing proposed network with state-of-the-art medical segmentation models indicate that ES-TransUNet maintains high segmentation accuracy while reducing model parameters and computational complexity significantly, and improves inference efficiency. It can be seen that ES-TransUNet is more satisfied the practical requirements in real-time medical image segmentation.

Key words: medical image segmentation, U-Net, lightweight, Transformer, skip connection, attention mechanism

中图分类号: