2D SECara-Net及3D U^(2)-Net模型用于检测MR血管造影中未破裂囊状颅内动脉瘤  

2D SECara-Net and 3D U^(2)-Net for detecting unruptured saccular intracranial aneurysms with MR angiography

作　　者：牛宗仁 马强 杜晶晶 任延德[1] 李梦杰 乔雅倩 唐岳山 高健波 NIU Zongren;MA Qiang;DU Jingjing;REN Yande;LI Mengjie;QIAO Yaqian;TANG Yueshan;GAO Jianbo(Department of Radiology,Affiliated Hospital of Qingdao University,Qingdao 266555,China;Department of Radiology,Zibo Maternal and Child Health Hospital,Zibo 255020,China;Department of Radiology,the First People's Hospital of Shizuishan City,Shizuishan 753200,China)

机构地区：[1]青岛大学附属医院放射科,山东青岛266555 [2]淄博市妇幼保健院影像科,山东淄博255020 [3]石嘴山市第一人民医院放射科,宁夏石嘴山753200

出　　处：《中国医学影像技术》2025年第2期245-249,共5页Chinese Journal of Medical Imaging Technology

摘　　要：目的观察基于2D最大密度投影(MIP)及3D时间飞跃法MR血管造影(3D TOF-MRA)图构建的2D SECara-Net、3D U^(2)-Net模型及其联合用于MRA检测未破裂囊状颅内动脉瘤(USIA)的价值。方法回顾性收集973例单发USIA及300名健康体检者,按7∶3比例划分训练集(n=923,含723例USIA及200名健康人)与测试集(n=350,含250例USIA及100名健康人)。分别将经预处理的训练集3D TOF-MRA及所获2D-MIP图导入3D U^(2)-Net及2D SECara-Net模型进行训练并调整参数;于测试集评估各模型及其联合检测USIA的效能。结果2D SECara-Net模型检测测试集USIA的敏感度为78.80%(197/250)、特异度95.00%(95/100)、准确率为83.43%(292/350),3D U^(2)-Net模型分别为82.80%(207/250)、86.00%(86/100)及83.71%(293/350),2D SECara-Net模型检测USIA的特异度高于3D U^(2)-Net模型(P=0.030)而敏感度及准确率差异均无统计学意义(P均>0.05);二者联合检测的敏感度为91.20%(228/250)、特异度为99.00%(99/100)、准确度为93.43%(327/350),其特异度高于3D U^(2)-Net模型(P<0.05)而敏感度及准确率均高于各单一模型(P均<0.05)。结论2D SECara Net与3D U^(2)-Net模型用于MRA检测USIA的敏感度及准确率相当;二者联合可提高检测效能。Objective To observe the value of 2D SECara-Net and 3D U^(2)-Net models constructed based on 2D maximal intensity projection(MIP)and 3D time-of-flight MR angiography(3D TOF-MRA)images,respectively,also of their combination for MRA detecting unruptured saccular intracranial aneurysms(USIA).Methods Totally 973 patients with single USIA and 300 subjects who underwent healthy physical examination were retrospectively collected and divided into training set(n=923,containing 723 cases of USIA and 200 healthy subjects)and test set(n=350,containing 250 cases of USIA and 100 healthy subjects)at the ratio of 7∶3.Pre-processed 3D TOF-MRA and the obtained 2D-MIP images in training set were imported into 3D U^(2)-Net and 2D SECara-Net models for training and adjusting parameters,respectively.The efficiency of 2 models and their combination for detecting USIA were evaluated.Results The sensitivity,specificity and accuracy of 2D SECara-Net model for detecting USIA in test set was 78.80%(197/250),95.00%(95/100)and 83.43%(292/350),of 3D U^(2)-Net model was 82.80%(207/250),86.00%(86/100)and 83.71%(293/350),respectively.The specificity of 2D SECara-Net model was higher than that of 3D U^(2)-Net model(P=0.030),while no significant difference of sensitivity nor accuracy was found between 2 models(both P>0.05).The specificity of the combination of the 2 models was 99.00%(99/100),higher than that of 3D U^(2)-Net model(P<0.05),and the sensitivity and accuracy of the combination was 91.20%(228/250)and 93.43%(327/350),respectivelty,both higher than those of 2 single models(all P<0.05).Conclusion 2D SECara-Net and 3D U^(2)-Net models had similar,sensitivity and accuracy for MRA detecting USIA.Combination of them could improve the detecting efficacy.

关 键 词：颅内动脉瘤 磁共振血管造影术 人工智能 

分 类 号：R543.5[医药卫生—心血管疾病] R445.2[医药卫生—内科学]