用于肺部肿瘤图像分割的跨模态多编码混合注意力U-Net  

Cross-modality Multi-encoder Hybrid Attention U-Net for Lung Tumors Images Segmentation

作　　者：周涛 董雅丽 刘珊 陆惠玲[3] 马宗军[4] 侯森宝 邱实[5] ZHOU Tao;DONG Yali;LIU Shan;LU Huiling;MA Zongjun;HOU Senbao;QIU Shi(School of Computer Science and Technology,North Minzu University,Yinchuan 750021,China;The Key Laboratory of Images&Graphics Intelligent Processing of State Ethnic Affairs Commission,North Minzu University,Yinchuan 750021,China;School of Science,Ningxia Medical University,Yinchuan 750004,China;Department of Orthopedics,Ningxia Medical University General Hospital,Yinchuan 750004,China;Xi'an Institute of Optics and Precision Mechanics,Chinese Academy of Sciences,Xi'an 710119,China)

机构地区：[1]北方民族大学计算机科学与工程学院,银川750021 [2]北方民族大学图像图形智能处理国家民委重点实验室,银川750021 [3]宁夏医科大学理学院,银川750004 [4]宁夏医科大学总医院骨科,银川750004 [5]中国科学院西安光学精密机械研究所,西安710119

出　　处：《光子学报》2022年第4期368-384,共17页Acta Photonica Sinica

基　　金：国家自然科学基金(No.62062003);宁夏自治区重点研发计划(No.2020BEB04022);北方民族大学引进人才科研启动项目(No.2020KYQD08);宁夏自然科学基金(No.2022AAC03149)。

摘　　要：针对医学图像的肺部肿瘤分割中病灶和周围组织的对比度低、边缘模糊、肿瘤和正常组织粘连、病灶和背景分布不均衡等问题,提出跨模态多编码混合注意力机制模型分割肺部病灶,用多种模态医学图像辅助分割病灶。首先设计了三编码器提取多模态医学图像的病灶特征,解决单模态医学影像的病灶特征提取能力不足的问题;然后针对网络通道维度冗余和对复杂病灶的空间感知能力不高的问题,在网络跳跃连接中加入混合注意力机制;最后对网络解码路径不同的尺度特征使用多尺度特征聚合块充分利用各个尺度特征。在临床多模态医学图像数据集上验证算法的有效性,对比实验结果表明所提模型对于肺部病灶分割的戴斯相似系数、召回率、体积重叠误差和相对体积差异分别为96.4%、97.27%、93.0%、93.06%。对于病灶形状复杂,病灶和正常组织粘连的情况,分割精度得到有效提升。The lung lesions segmentation in medical imaging is an important task. However,there are still some challenges. The lesions delineation relies on manual segmentation by experienced clinicians,which is time-consuming and labor-intensive due to the complex anatomical structure of the human body;Lung tumor images have the characteristics of low contrast,different size and shape of the lesions,and variable location of the lesions,and are characterized by unbalanced data distribution. U-Net can segment lesions under a small number of datasets and has been widely used in medical image segmentation of lesions and organs. However,U-Net has the following three problems. First,U-Net uses uniform parameters for each feature map. For lesions of different sizes and complex shapes,the network may have poor spatial perception,which leads to the decline of segmentation performance. Second,U-Net channel dimension doubles with the number of down-sampling,and the feature map of the encoder layer is concatenated to the decoding layer through skip connection. However,in the segmentation task,the importance of different channels to the segmentation task is different. Third,most of the current multi-encoder segmentation networks extract the features of the single-modal target slice and their continuous slices to improve the network segmentation performance,but ignore the ability of different modal medical images to express the characteristics of the lesion. To solve the above problems,this paper proposes the MEAU-Net network to extract complementary features of multi-modals images. First,for the unbalanced data distribution,the Hough transform is used to detect the line of the lung Computed Tomography(CT)image marked by the doctor to obtain the region of interest,and cropped image size from 356 pixel×356 pixel to 50 pixel×50 pixel. Then,for the low contrast of medical image,use exposure fusion image contrast enhancement method improves the contrast between lesion and the background of lung CT image. To extract the features of lesions

关 键 词：深度学习 医学图像分割 多模态医学图像 U-Net 肺癌 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]