基于注意力机制的U-Net脑脊液细胞分割  

U-Net CSF Cells Segmentation Based on Attention Mechanism

作　　者：代茵[1,2] 刘维宾 董昕阳 宋雨朦 DAI Yin;LIU Wei-bin;DONG Xin-yang;SONG Yu-meng(College of Medicine and Biological Information Engineering,Northeastern University,Shenyang 110169,China;Engineering Center on Medical Imaging and Intelligent Analysis,Ministry of Education,Northeastern University,Shenyang 110169,China;School of Computing,University of York,Yorkshire YO105DD,UK)

机构地区：[1]东北大学医学与生物信息工程学院,辽宁沈阳110169 [2]东北大学教育部医学影像与智能分析工程中心,辽宁沈阳110169 [3]约克大学计算机学院,英国约克郡YO105DD

出　　处：《东北大学学报（自然科学版）》2022年第7期944-950,共7页Journal of Northeastern University(Natural Science)

基　　金：国家自然科学基金青年基金资助项目(61902058);中央高校基本科研业务费专项资金资助项目(N2019002).

摘　　要：为解决脑脊液病理图像中部分细胞膜较为模糊,与图像背景难以区分的问题,采用了基于注意力机制的U-Net深度学习方法对脑脊液病理图像做全自动分割.在深度学习网络中加入注意力机制对细胞进行定位,抑制无关信息,提高语义的特征表达,提高对细胞整体分割的精确性.通过镜像、旋转等操作对数据集进行扩充预处理.采用VGG16预训练模型进行迁移学习,交叉熵与Dice损失相结合作为损失函数,分别在脑脊液临床图像与公开数据集2018 Data Science Bowl上进行验证;并与Otsu,PSPnet,Segnet,DeeplabV3+,U-Net进行对比,结果表明,本文方法在各项指标上均优于其他分割方法.In order to solve the problem that part of the cell membrane in the pathological images of CSF(cerebrospinal fluid)is blurred and this is difficult to be distinguished from the image background.The U-Net based on attention mechanism is proposed to segment pathological images of CSF automatically.Attention mechanism is added to deep learning network to locate cells,suppress irrelevant information,improve semantic feature expression,and further improve the accuracy of cell segmentation.The datasets are preprocessed by mirroring and rotation.VGG16 pre-training model is used for transfer learning.Cross entropy is combined with Dice loss as Loss function which is validated in CSF clinical images and open dataset 2018 Data Science Bowl and compared with Otsu,PSPnet,Segnet,DeeplabV3+,U-Net.The results show that the proposed method is superior to other segmentation methods in all indexes.

关 键 词：脑脊液检测 细胞分割 注意力机制 深度学习 U-Net模型 

分 类 号：TP391.5[自动化与计算机技术—计算机应用技术]