基于多尺度特征融合和网格注意力机制的三维肝脏影像分割方法  

3D liver image segmentation method based on multi-scale feature fusion and grid attention mechanism

作　　者：郑帅 张晓龙[1,2,3] 邓鹤 任宏伟[4] ZHENG Shuai;ZHANG Xiaolong;DENG He;REN Hongwei(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan Hubei 430065,China;Institute of Big Data Science and Engineering,Wuhan University of Science and Technology,Wuhan Hubei 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System(Wuhan University of Science and Technology),Wuhan Hubei 430065,China;Tianyou Hospital Affiliated to Wuhan University of Science and Technology,Wuhan Hubei 430064,China)

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

出　　处：《计算机应用》2023年第7期2303-2310,共8页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61972299,62071456)。

摘　　要：在计算机断层扫描(CT)和磁共振成像(MRI)的影像中肝脏与邻近脏器的灰度值相似性都比较高,为自动精确地分割肝脏,提出一种基于多尺度特征融合和网格注意力机制的三维肝脏影像分割方法MAGNet(Multi-scale feature fusion And Grid attention mechanism Network)。首先,通过注意力引导连接模块来连接高层特征和低层特征以提取出重要的上下文信息,并且在注意力引导连接模块中引入网格注意力机制来关注感兴趣的分割区域;然后,通过在单个特征图中按通道数进行分层连接形成多尺度特征融合模块,并用该模块替换基础卷积块以获取多尺度语义信息;最后,利用深度监督机制解决梯度消失、梯度爆炸和收敛过慢等问题。实验结果表明:在3DIRCADb数据集上,与U3-Net+DC方法相比,MAGNet在Dice相似系数(DSC)指标上提升了0.10个百分点,在相对体积差(RVD)指标上降低了1.97个百分点;在Sliver07数据集上,与CANet方法相比,MAGNet在DSC指标上提升了0.30个百分点,在体素重叠误差(VOE)指标上降低了0.68个百分点,在平均对称表面距离(ASD)和对称位置表面距离的均方根(RMSD)指标上分别降低了0.03 mm和0.22 mm;在某医院肝脏MRI数据集上,MAGNet在所有指标上也均具有良好的结果。另外,将MAGNet应用于3DIRCADb数据集和某医院肝脏MRI数据集进行混合形成的数据集,也取得了非常有竞争力的分割效果。Due to the high similarity of gray values among liver and adjacent organs in Computed Tomography(CT)and Magnetic Resonance Imaging(MRI)images,a 3D liver image segmentation method based on multi-scale feature fusion and grid attention mechanism,namely MAGNet(Multi-scale feature fusion And Grid attention mechanism Network),was proposed to segment liver automatically and accurately.Firstly,high-level features and low-level features were connected by the attention-guided concatenation module to extract important context information,and the grid attention mechanism was introduced in the attention-guided concatenation module to focus on the segmentation region of interest.Then,the multiscale feature fusion module was formed by the layered connection in a single feature map according to the number of channels,and this module was used to replace the basic convolutional block to obtain multi-scale semantic information.Finally,the deep supervision mechanism was utilized to solve the problems of vanishing gradient,exploding gradient and slow convergence.Experimental results show that on 3DIRCADb dataset,compared with the U3-Net+DC method,MAGNet improves the Dice Similarity Coefficient(DSC)metric by 0.10 percentage points and reduces the Relative Volume Difference(RVD)metric by 1.97 percentage points;on Sliver07 dataset,compared with the CANet method,MAGNet improves the DSC metrics by 0.30 percentage points,reduces Volumetric Overlap Error(VOE)metrics by 0.68 percentage points,and reduces the Average Symmetric Surface Distance(ASD)and Root Mean Square Symmetric Surface Distance(RMSD)metrics 0.03 mm and 0.22 mm respectively;on the liver MRI dataset of a hospital,MAGNet also has good results on all metrics.Besides,MAGNet was applied to a mixed dataset of 3DIRCADb dataset and the hospital liver MRI dataset above,and a competitive segmentation result was also achieved.

关 键 词：三维肝脏医疗影像 语义分割 深度学习 多尺度特征融合 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]