基于改进TransUNet模型的脑肿瘤图像分割方法研究  

Research on Brain Tumor Image Segmentation Method Based on Improved TransUNet Model

作　　者：朱玉婷 袁晓[1] ZHU Yuting;YUAN Xiao(College of Electronic and Information,Sichuan University,Chengdu,Sichuan 610064,China)

机构地区：[1]四川大学电子信息学院,四川成都610064

出　　处：《计算技术与自动化》2024年第2期98-104,共7页Computing Technology and Automation

摘　　要：针对肿瘤细胞图像与正常组织图像之间具有强相似性、边界模糊以及染色变化大等特点,提出了基于TransUNet网络的优化改进分割模型。此分割模型在以TransUNet为主干网络的基础上于编码器部分引入注意力机制,抑制不相关的部分以突显深层特征的语义信息。同时,改变上采样过程中的融合方式,引入BiFusion模块进行选择性地融合,从而使特征数据能够保留更多高分辨率细节信息。该分割模型在Kaggle脑部低级别胶质瘤数据集上验证。实验结果表明,改进后算法的均交并比,召回率和平均精度均值分别为:97.31%,99.91%和98.72%,与目前医学图像分割的主流方法相比具有更优的性能。In view of the strong similarity between tumor cell image and normal tissue image,fuzzy boundary and large staining change,an improved optimization model based on TransUNet network is proposed.Based on TransUNet as the backbone network,this segmentation model introduces the attention mechanism into the encoder part to suppress the irrelevant part to highlight the semantic information of the deep features.At the same time,the fusion method in the up-sampling process is changed,and the BiFusion module is introduced for selective fusion,so that the feature data can retain more high-resolution details.The model is verified on the Kaggle brain low-grade glioma data set.The experimental results show that the M IoU,R and m AP of the improved algorithm are 97.31%,99.91%and 98.72%respectively,which has better performance compared with the current mainstream methods of medical image segmentation.

关 键 词：脑肿瘤 医学图像分割 注意力机制 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]