融合卷积和Transformer的腹部多器官分割网络  

Abdominal Multi Organ Segmentation Network Combining Convolution and Transformer

作　　者：杨萍 陈立伟[1] 王庆凤[1] 周莹[2] YANG Ping;CHEN Li-wei;WANG Qing-feng;ZHOU Ying(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621000,China;Radiology Department,Mianyang Center Hospital,Mianyang 621000,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621000 [2]绵阳市中心医院放射科,四川绵阳621000

出　　处：《计算机技术与发展》2024年第9期47-54,共8页Computer Technology and Development

基　　金：四川省自然科学基金项目(2022NSFSC0940,2022NSFSC0894);西南科技大学博士基金项目(19zx7143,20zx7137)。

摘　　要：腹部多器官分割在计算机辅助诊断中起着至关重要的作用,具有重要的研究价值。但由于腹部多器官边界模糊、背景复杂以及形状大小多变,使这项任务极具挑战性。为此,提出了一种融合卷积和Transformer的腹部多器官分割网络TCMSUnet。首先,在特征提取阶段设计了多尺度引导融合模块(GFM),利用高层特征提取的显著语义信息来引导低层特征以增强相邻特征的语义一致性,从而促进不同尺度特征的融合;随后设计了全局局部增强模块(GLE),通过空洞卷积和Transformer块结合来增强模型对全局局部上下文信息的提取,使模型在建立长距离依赖关系的同时加强特征的局部关联性;最后,在解码器部分引入多阶段损失聚合结构以加快模型的收敛并优化模型的性能。在Synapse数据集上评估了模型的性能,其平均Dice相似系数(DSC)为81.20%。实验结果表明,所提方法整体性能优于多种比较网络,并对形状大小多变的器官有更好的分割效果。Abdominal multi organ segmentation plays a crucial role in computer-aided diagnosis and has significant research value.However,due to the blurred boundaries of multiple organs in the abdomen,complex backgrounds,and variable shapes and sizes,this task is extremely challenging.To this end,TCMSUnet,a new abdominal multi organ segmentation network that integrates convolution and Transformer is proposed.Firstly,a multi-scale guided fusion module(GFM)was designed in the feature extraction stage,which utilizes the significant semantic information extracted from high-level features to guide low-level features and enhance the semantic consistency of adjacent features,thereby promoting the fusion of features at different scales.Subsequently,a global local enhancement module(GLE)was designed to enhance the model’s extraction of global and local contextual information through a combination of dilated convolution and Transformer blocks,enabling the model to establish long-range dependencies while enhancing local correlations of features.Finally,a multi-stage loss aggregation structure was introduced in the decoder section to accelerate the convergence of the model and optimize its performance.The performance of the model was evaluated on the Synapse dataset,with an average Dice similarity coefficient(DSC)of 81.20%.The experimental results show that the proposed method outperforms multiple comparison networks in overall performance and has better segmentation performance for organs with variable shapes and sizes.

关 键 词：医学图像分割 特征融合 多尺度 空洞卷积 TRANSFORMER 多器官 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]