基于SF-TransUNet的腹部多器官图像分割方法  

The Abdominal Multi-organ Image Segmentation Method Based on SF-TransUNet

作　　者：郭雨婷 于瓅[1] GUO Yuting;YU Li(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《湖北民族大学学报(自然科学版)》2025年第1期94-100,共7页Journal of Hubei Minzu University:Natural Science Edition

基　　金：安徽省重点研究与开发计划项目(202104d07020010)。

摘　　要：针对腹部多器官图像分割过程中小器官图像分割精度较低和边界分割不准确的问题,提出了基于变换器U形网络(Transformer U-shaped network, TransUNet)的融合分割TransUNet(segmentation fusion TransUNet, SF-TransUNet)改进模型,以增强小器官图像分割精度。在TransUNet跳跃连接中加入用于增强纹理信息的改进位置注意力模块(position attention module, PAM),并在解码器中引入混洗注意力(shuffle attention, SA)模块融合高低层特征,提高小器官图像细节捕捉能力,设计连通域分析(connected component analysis, CCA)模块作为后处理步骤,有效提升边缘分割能力。在Synapse数据集上验证SF-TransUNet模型性能,结果显示其平均戴斯相似系数(Dice similarity coefficient, DSC)比TransUNet模型提升了2.69个百分点,95%豪斯多夫距离(95%Hausdorff distance, HD95)下降了17.26 mm;在小器官图像分割上,胆囊、右肾和胰腺的分割精度分别提高了9.22、4.76、4.49个百分点。结果表明,SF-TransUNet模型不仅明显提升了腹部多器官图像分割的总体精度,而且在小器官图像分割中能表现出更佳的特征表达与细节保留能力。To address the challenge of low image segmentation accuracy for small organs in abdominal multi-organ image segmentation,an improved model based on the Transformer U-shaped network(TransUNet)called segmentation fusion TransUNet(SF-TransUNet)was proposed to enhance the image segmentation accuracy of small organs.The improved position attention module(PAM)for texture information enhancement was introduced into the skip connections of TransUNet,and a high-and low-level feature fusion shuffle attention(SA)module was added in the decoder to improve the capture ability of fine details for small organs image.Additionally,connected component analysis(CCA)module was designed as a post-processing step to effectively enhance edge segmentation capabilities.The experiments on the Synapse dataset validated the performance of SF-TransUNet model,with results showing that the average Dice similarity coefficient(DSC)had an increase of 2.69 percentage points compared to the TransUNet model,and the 95%Hausdorff distance(HD95)had a decrease of 17.26 mm.For small organs,image segmentation accuracy for the gallbladder,right kidney and pancreas improved by 9.22,4.76 and 4.49 percentage points,respectively.The findings demonstrated that SF-TransUNet model not only enhanced the overall accuracy of abdominal multi-organ image segmentation significantly but also exhibited superior feature representation and detail retention for small organ image segmentation.

关 键 词：SF-TransUNet 多器官图像分割 小器官图像分割 纹理信息增强 混洗注意力 连通域分析 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] R319.1[自动化与计算机技术—计算机科学与技术]