基于多尺度空间Transformer的肝脏分割方法  

Liver Segmentation Method Based on Multi-scale Spatial Transformer

作　　者：丁厚林 张晓龙[1,2,3] 林晓丽[1,2,3] 邓鹤 任宏伟[4] DING Hou-lin;ZHANG Xiao-long;LIN Xiao-li;DENG He;REN Hong-wei(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Institute of Big Data Science and Engineering,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Key Laboratory of Intelligent Information Processing and Real-time Industrial System,Wuhan University of Science and Technology,Wuhan 430065,China;Tianyou Hospital Affiliated to Wuhan University of Science and Technology,Wuhan 430064,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]武汉科技大学大数据科学与工程研究院,湖北武汉430065 [3]武汉科技大学智能信息处理与实时工业系统湖北省重点实验室,湖北武汉430065 [4]武汉科技大学附属天佑医院,湖北武汉430064

出　　处：《计算机技术与发展》2025年第2期1-8,共8页Computer Technology and Development

基　　金：国家自然科学基金项目(61972299,62071456)。

摘　　要：肝脏器官尺度多样且与周围器官高度相似,很难从腹部计算机影像中准确分割出肝脏区域,现有的很多方法将CNN和Transformer相结合以得到图像局部和全局特征依赖关系,从而取得了更好的性能。然而,简单的组合方法忽视了图像分割中多尺度特征融合和注意力机制的重要性,没有很好地解决肝脏分割问题。该文提出了一种用于肝脏分割的多尺度空间Transformer与交叉自注意机制的三维肝脏影像分割方法。该方法首先采用CNN和Transformer相结合的方式逐步提取不同尺度的特征信息使网络对肝脏及其周围组织的识别更加准确;接着利用多尺度空间Transformer对不同层次和尺度特征的图像在空间维度上融合,提高了网络对肝脏边缘的定位能力;最后在解码器中设计了交叉自注意引导融合模块减少噪声等不相关信息带来的干扰,提高分割质量。在LiTS、CHAOS、Sliver07和某医院MRI数据集上进行了对比和消融实验,实验结果表明,该方法相较于当前的主流网络具有更好的分割性能和临床应用前景。The liver organs have diverse scales and are highly similar to surrounding organs,making it difficult to accurately segment the liver region from abdominal computer images.Many existing methods combine CNN and Transformer to obtain local and global feature dependencies of the image,achieving better performance.However,simple combination methods have overlooked the importance of multi-scale feature fusion and attention mechanisms in image segmentation,and have not effectively solved the problem of liver segmentation.We propose a 3D liver image segmentation method using multi-scale spatial Transformer and cross self-attention mechanism for liver segmentation.The method first uses a combination of CNN and Transformer to gradually extract feature information of different scales,making the network's recognition of the liver and its surrounding tissues more accurate.Then,the multi-scale spatial Transformer is used to fuse images with different levels and scales in the spatial dimension,improving the network's ability to locate the edges of the liver.Finally,a cross self attention guided fusion module is designed in the decoder to reduce interference caused by irrelevant information such as noise and improve segmentation quality.The proposed method is compared and subjected to ablation experiments on LiTS,CHAOS,Sliver07,and a hospital MRI dataset.The experimental results show that the proposed method has higher segmentation performance and clinical application prospects compared to current mainstream networks.

关 键 词：三维肝脏影像分割 深度学习 交叉自注意机制 多尺度空间Transformer 多尺度特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]