基于扩散加权成像的密集卷积神经网络模型预测急性缺血性脑卒中TOAST病因分型的研究  

Diffusion-weighted imaging-based DenseNet model for prediction of TOAST etiological typing in acute ischemic stroke

作　　者：帕哈提·吐逊江 赵伟 罕迦尔别克·库锟 徐蕊 常一凡 艾尼卡尔江·艾合麦提 许峥 王云玲 Tuxunjiang Pahati;Zhao Wei;Kukun Hanjiaerbieke;Xu Rui;Chang Yifan;Aihemaiti Ainikaerjiang;Xu Zheng;Wang Yunling(Department of Radiology,First Affiliated Hospital of Xinjiang Medical University,Urumqi 830054,China;Shukun Technology Co.,Ltd,Beijing 102200,China)

机构地区：[1]新疆医科大学第一附属医院影像中心,乌鲁木齐830054 [2]数坤科技股份有限公司,北京102200

出　　处：《中华放射学杂志》2024年第10期1015-1020,共6页Chinese Journal of Radiology

基　　金：中央引导地方科技发展资金项目(ZYYD2023D02);“天山英才”科技创新领军人才项目(2023TSYCLJ0027)。

摘　　要：目的探讨基于扩散加权成像(DWI)的深度学习模型快速识别急性缺血性脑卒中(AIS)患者TOAST病因分型的价值。方法该研究为横断面研究。回顾性分析2023年3月至2024年2月新疆医科大学第一附属医院收治的504例AIS患者的影像和临床资料。根据TOAST病因分型分为大动脉粥样硬化型252例、小动脉闭塞型252例。504例患者按6∶2∶2的比例采用分层随机法分为训练集(302例)、验证集(101例)和测试集(101例)。所有患者均接受DWI,采用DenseNet网络框架,通过优化不同层数的模型配置分别构建不同层数(121层、169层、201层)的DenseNet模型(DenseNet169模型、DenseNet121模型、DenseNet201模型)。通过使用数据增强、Adam优化器及交叉熵损失函数方法提高模型收敛速度、鲁棒性及平衡正负样本不平衡问题。采用独立样本t检验或χ^(2)检验比较大动脉粥样硬化型和小动脉闭塞型AIS患者临床资料。绘制受试者操作特征曲线和曲线下面积(AUC)评估各模型鉴别大动脉粥样硬化型和小动脉闭塞型AIS患者的效能。结果大动脉粥样硬化型与小动脉闭塞型AIS患者中年龄、入院美国国立卫生院卒中量表评分、大血管狭窄/闭塞差异具有统计学意义(P<0.05)。3种DenseNet模型在训练集、验证集及测试集中均显示出较好的区分大动脉粥样硬化型和小动脉闭塞型的能力。在测试集中,DenseNet201模型鉴别大动脉粥样硬化型与小动脉闭塞型AIS患者的AUC、灵敏度、准确度及F1分数值(分别为0.826、0.902、0.743、0.780)均高于DenseNet121(分别为0.801、0.647、0.723、0.702)、DenseNet169模型(0.778、0.882、0.733、0.769)的AUC、灵敏度、准确度及F1分数值。结论基于深度学习模型能够实现对AIS患者自动化TOAST病因分型,且DenseNet201模型的分类效能最优且性能更稳定。Objective To investigate the value of a deep learning model based on diffusion-weighted imaging(DWI)in quick identification of the TOAST etiology classification in patients with acute ischemic stroke(AIS).Methods In this cross-sectional study,imaging and clinical data of 504 patients with AIS admitted to the First Affiliated Hospital of Xinjiang Medical University from March 2023 to February 2024 were retrospectively reviewed.Using the TOAST etiology classification,there were 252 large artery atherosclerosis type and 252 small-artery occlusion type.The 504 cases were divided into a training set(n=302),a validation set(n=101)and a test set(n=101)using stratified randomization in the ratio of 6∶2∶2.All cases had DWI data.A DenseNet network framework was used to construct DenseNet models by optimizing the model configurations of different layers.Three DenseNet models with different layers(121,169,201)were constructed,named DenseNet169 model,DenseNet121 model,and DenseNet201 model.The data enhancement,Adam optimizer and cross-entropy loss function methods were used to improve the convergence speed and robustness of the model,and to balance the positive and negative sample imbalance problem.Independent sample t-test or χ^(2) was used to compare the clinical data of patients with large artery atherosclerosis type and small-artery occlusion type AIS.Receiver operating characteristic curves and area under the curve(AUC)were performed to evaluate the efficacy of each model in identification of patients with large artery atherosclerosis type and small-artery occlusion type AIS.Results There were statistically significant differences in age,National Institutes of Health Stroke Scale score at admission,and stenosis or occlusion of large vessels between patients with large artery atherosclerosis type and small-artery occlusion(all P<0.05).In the test set,the AUC,sensitivity,accuracy,and F1 score values of the DenseNet201 model for discriminating patients with large artery atherosclerosis type AIS and small-artery occlusi

关 键 词：卒中 深度学习 扩散加权成像 病因分型 

分 类 号：R445.2[医药卫生—影像医学与核医学] R743.3[医药卫生—诊断学]