基于知识引导的缺血性脑卒中梗死区分割方法  

Knowledge-guided infarct segmentation of ischemic stroke

作　　者：顾正宇 赖菲菲 耿辰[3] 王希明[4] 戴亚康 GU Zhengyu;LAI Feifei;GENG Chen;WANG Ximing;DAI Yakang(School of Medical Imaging,Xuzhou Medical University,Xuzhou 221004,China;Department of Radiology,Wuxi Mental Health Center,Wuxi 214151,China;Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Sciences,Suzhou 215163,China;Department of Radiology,The First Affiliated Hospital of Soochow University,Suzhou 215006,China)

机构地区：[1]徐州医科大学医学影像学院,江苏徐州221004 [2]无锡市精神卫生中心放射科,江苏无锡214151 [3]中国科学院苏州生物医学工程技术研究所,江苏苏州215163 [4]苏州大学附属第一医院放射科,江苏苏州215006

出　　处：《浙江大学学报(工学版)》2025年第4期814-820,共7页Journal of Zhejiang University:Engineering Science

基　　金：国家自然科学基金资助项目(81971685,62441114);江苏省前沿引领技术基础研究项目(BK20192004);山东省自然科学基金资助项目(ZR2022QF093);苏州科技计划项目(SKY2022151);浙江省医药卫生科技计划项目(2022KY1426).

摘　　要：针对缺血性脑卒中梗死区在医学影像上显示出低密度特征,提出基于阈值分割和加权滤波的梗死概率图生成方法.通过自适应参数的阈值分割找出低密度区,由多尺度自定义权重的滤波器计算二值图,获得梗死概率图.当概率图引导网络参数学习时,通过降低低概率区域的权重来提高所提方法的分割准确度.使用梗死概率图引导U-Net、DRINet和DeepLabV3+,相比未使用梗死概率图引导的模型,Dice系数分别提升了0.0466,0.0418和0.0363,交并比(IoU)分别提升了0.0322、0.0440和0.0356.统计结果表明,梗死概率图引导的网络对急性期数据的Dice系数有提升作用,对亚急性数据分割结果影响不大.所提方法为自动分割急性缺血性脑卒中梗死区提供了可行方案.Ischemic stroke infarcts show low-density features on imaging.Based on threshold segmentation and weighted filtering for the features,an infarct probability map generation method was proposed.The low-density region was identified by threshold segmentation with adaptive parameters,and the infarct probability map was obtained by calculating the binary map through the filter with customized weights at multiscale.The weights of the low-probability regions were reduced when the probability map guided the net parameters learning,thus improving the segmentation accuracy of the proposed method.The probability map was used to guide U-Net,DRINet and DeepLabV3+,the Dice coefficient was increased by 0.0466,0.0418 and 0.0363,and the intersection over union(IoU)was increased by 0.0322,0.0440 and 0.0356,respectively,compared to the models not guided with infarct probability maps.Statistical results show that the infarct probability map-guided network has an enhancement effect on the Dice coefficient for acute-phase data and has little effect on the segmentation results for sub-acute data.The proposed method provides a feasible solution for automatic segmentation of the acute ischemic stroke infarct.

关 键 词：深度学习 知识引导 缺血性卒中 脑卒中梗死区 语义分割 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]