摘要
基于卷积神经网络的方法在医学图像分割任务中取得了显著成果,但该方法固有的归纳偏置使其不能很好地学习全局和长距离的语义信息交互,而Transformer的优势是关注全局信息,两者可以优势互补。因此提出一种针对分割边缘利用Swin Transformer融合边缘感知的医学图像分割网络。设计基于上下文金字塔的边缘感知模块,用于融合全局的多尺度的上下文信息,针对边缘和角落等局部特征,利用浅层深度主干的特征产生丰富的边缘特征,因此提出的边缘感知模块可以尽可能多地产生边缘特征。在腹部多器官分割任务和心脏分割数据集的实验结果表明,该方法在各项指标中都有所提高。
The method based on convolution neural network has achieved remarkable results in the medical image segmentation task,but the inherent inductive bias of this method makes it unable to learn the global and long-distance semantic information interaction well,while the advantage of Transformer is to focus on the global information,and the two can complement each other.Therefore,a medical image segmentation network based on Swin Transformer fusion edge perception was proposed.The edge sensing module based on the context pyramid was designed to fuse the global multi-scale context information,mainly for the local features such as edges and corners,and the features of the shallow depth backbone were used to generate edge features.Therefore,the proposed edge sensing module could extract as many shallow features as possible to generate edge features.Results of experiments on abdominal multi organ segmentation task and heart segmentation dataset show that the method improves all indicators.
作者
叶晋豫
李娇
邓红霞
张瑞欣
李海芳
YE Jin-yu;LI Jiao;DENG Hong-xia;ZHANG Rui-xin;LI Hai-fang(College of Information and Computer,Taiyuan University of Technology,Jinzhong 030600,China)
出处
《计算机工程与设计》
北大核心
2024年第4期1149-1156,共8页
Computer Engineering and Design
基金
浙江大学CAD&CG国家重点实验室2022年开放基金项目(A2221)
山西省中央引导地方科技发展基金项目(YDZJSX2022A016)
国家自然科学基金项目(61976150)
山西省自然科学基金项目(201901D111091)。
关键词
医学图像分割
移动窗口变形器
多头自注意力
边缘感知模块
上下文金字塔
多尺度特征
深度学习网络
medical image segmentation
Swin Transformer
multiple self-attention
edge awareness module
context pyramid fusion network
multiscale feature
deep learning network
