摘要
目的探讨基于深度学习的快速磁敏感加权成像(SWI)评估急性缺血性卒中(AIS)的价值。方法回顾性分析2019年1月至2021年1月在解放军总医院第一医学中心接受MR检查并行SWI序列扫描且发病24 h内的AIS患者118例,其中男75例、女43例,年龄23~100(66±14)岁。采用MATLAB的randperm函数将118例患者以8∶2的比例分为训练集96例,测试集22例。另外收集MR-STARS研究招募的同一中心的47例AIS患者作为外部验证集,其中男38例、女9例,年龄16~75(58±12)岁。将SWI图像和滤波相位图像合并为复值图像作为全采样参考图像。对全采样参考图像进行回顾性欠采样以模拟实际欠采样过程,获得欠采样SWI图像,欠采样倍数为5倍,可节省80%的扫描时间;然后基于复值卷积神经网络(ComplexNet)的深度学习模型重建快速SWI数据。采用组内相关系数(ICC)或Kappa检验比较全采样SWI和基于ComplexNet的快速SWI的图像质量评分一致性及检出AIS患者磁敏感血栓征(SVS)、微出血、深部髓质静脉(DMVs)不对称的一致性。结果测试集中全采样SWI图像质量的评分为(4.5±0.6)分,基于ComplexNet的快速SWI图像质量评分为(4.6±0.7),两者一致性较好(ICC=0.86,P<0.05);全采样SWI与基于ComplexNet的快速SWI检出AIS患者SVS(Kappa=0.79,P<0.05)、微出血(Kappa=0.86,P<0.05)、DMVs不对称(Kappa=0.82,P<0.05)一致性较好。外部验证集中,全采样SWI图像质量的评分为(4.1±1.0)分,基于ComplexNet的快速SWI图像质量的评分为(4.0±0.9)分,两者一致性较好(ICC=0.97,P<0.05);全采样SWI与基于ComplexNet的快速SWI检出AIS患者SVS(Kappa=0.74,P<0.05)、微出血(Kappa=0.83,P<0.05)和DMVs不对称(Kappa=0.74,P<0.05)的一致性较好。结论深度学习技术可显著加快SWI速度,且基于ComplexNet的快速SWI与全采样SWI的图像质量及检出AIS征象的一致性好,可应用于AIS患者的影像学评估。
Objective To explore the value of fast susceptibility weighted imaging(SWI)generated by a deep learning model in assessment of acute ischemic stroke(AIS).Methods From January 2019 to January 2021,118 AIS patients[75 males and 43 females,aged 23-100(66±14)years]who underwent MR examination and SWI sequence scanning within 24 h of symptom onset in the First Medical Center of PLA General Hospital were retrospectively analyzed.MATLAB′s randperm function was used to divide 118 patients into a training set of 96 cases and a test set of 22 cases at a ratio of 8∶2.Fourty-seven AIS patients[38 males and 9 females,aged 16-75(58±12)years]from one center of a multicenter study were selected to build the external validation set.SWI image and filtered phase image were combined into complex value image as full sampling reference image.Undersampled SWI images were obtained by retrospective undersampling of reference fully sampled images,and the undersampling multiple was five times which could save 80%of the scanning time,then the complex-valued convolutional neural network(ComplexNet)was used to develop reconstruct fast SWI.Interclass correlation coefficient(ICC)or Kappa tests were used to compare the consistency of image quality and the diagnostic consistency for the presence of susceptibility vessel sign(SVS),cerebral microbleeds and asymmetry of cerebral deep medullary veins(DMVs)in AIS patient on fully sampled SWI and fast SWI based on ComplexNet.Results In test set,score of image quality was 4.5±0.6 for fully sampled SWI image and 4.6±0.7 for fast SWI based on ComplexNet,and coefficient was excellent(ICC=0.86,P<0.05).Full sampling SWI had good agreement with fast SWI based on ComplexNet in detecting SVS(Kappa=0.79,P<0.05),microbleeds(Kappa=0.86,P<0.05),and DMVs asymmetry(Kappa=0.82,P<0.05)in AIS patients.In the external validation set,score of image quality was 4.1±1.0 for fully sampled SWI image and 4.0±0.9 for fast SWI based on ComplexNet,and coefficient was excellent(ICC=0.97,P<0.05).Full sampling SWI had good agreement with fast SWI based on ComplexNet in detecting SVS(Kappa=0.74,P<0.05),microbleeds(Kappa=0.83,P<0.05),and DMVs asymmetry(Kappa=0.74,P<0.05)in AIS patients.Conclusions Deep learning techniques can significantly accelerate the speed of SWI,and the consistency of image quality and detected AIS signs between fast SWI based on ComplexNet and fully sampled SWI is good.The fast SWI based on ComplexNet can be applied to the radiographic assessment of clinical AIS patients.
作者
段祺
段曹辉
周世擎
吕晋浩
边祥兵
张德康
程焜
杨明亮
王雪扬
张汀阳
邢新博
田成林
娄昕
Duan Qi;Duan Caohui;Zhou Shiqing;Lyu Jinhao;Bian Xiangbing;Zhang Dekang;Cheng Kun;Yang Mingliang;Wang Xueyang;Zhang Tingyang;Xing Xinbo;Tian Chenglin;Lou Xin(Medical School of Chinese PLA,Beijing 100853,China;Department of Radiology,First Medical Center of PLA General Hospital,Beijing 100853,China;Department of Neurology,First Medical Center of PLA General Hospital,Beijing 100853,China)
出处
《中华放射学杂志》
CAS
CSCD
北大核心
2023年第1期34-40,共7页
Chinese Journal of Radiology
基金
国家自然科学基金(81825012,81730048,82151309,81901708)。
关键词
卒中
磁共振成像
磁敏感加权成像
深度学习
Stroke
Magnetic resonance imaging
Susceptibility weighted imaging
Deep learning