摘要
脑动脉瘤破裂造成的蛛网膜下腔出血致死致残率极高,借助深度学习网络辅助医生实现高效筛查具有重要意义.为提高基于时间飞跃法磁共振血管造影(Time of Flight-Magnetic Resonance Angiography,TOF-MRA)的脑动脉瘤自动检测的精度,本文基于模糊标签方式,提出一种基于变体3D U-Net和双分支通道注意力(Dual-branch Channel Attention,DCA)的深度神经网络DCAU-Net,DCA模块可以自适应地调整通道特征的响应,提高特征提取能力.首先对260例病例的TOF-MRA影像预处理,将数据集分为174例训练集、43例验证集和43例测试集,然后使用处理后的数据训练和验证DCAU-Net,测试集实验结果表明DCAU-Net可以达到90.69%的敏感度,0.83个/例的假阳性计数和0.52的阳性预测值,有望为动脉瘤筛查提供参考.
Subarachnoid hemorrhage caused by the rupture of cerebral aneurysms is extremely fatal and disabling.It’s imperative for radiologists to achieve efficient screening with the help of deep learning-based models.To improve the detection sensitivity of time of flight-magnetic resonance angiography(TOF-MRA)images,this study proposed a neural network named DCAU-Net which is based on fuzzy labels,3D U-Net variant,and dual-branch channel attention(DCA),and able to adaptively adjust the response of channel features to improve feature extraction capability.First,TOF-MRA images from 260 subjects were preprocessed,and the data were split into the training set(N=174),validation set(N=43)and testing set(N=43).Then the preprocessed data were used for training and validating DCAU-Net.The results show that DCAU-Net scores 90.69%of sensitivity,0.83 per case of false positive count and 0.52 of positive predicted value in the testing set,providing a promising tool for detecting cerebral aneurysms.
作者
陈萌
耿辰
李郁欣
耿道颖
鲍奕仿
戴亚康
CHEN Meng;GENG Chen;LI Yu-xin;GENG Dao-ying;BAO Yi-fang;DAI Ya-kang(School of Medical Imaging,Xuzhou Medical University,Xuzhou 221000,China;Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Sciences,Suzhou 215000,China;Department of Radiology,Huashan Hospital,Fudan University,Institute of Functional and Molecular Medical Imaging,Fudan University,Shanghai 200000,China)
出处
《波谱学杂志》
CAS
北大核心
2022年第3期267-277,共11页
Chinese Journal of Magnetic Resonance
基金
国家自然科学基金资助项目(61672236)
上海市科学技术委员会科技创新行动计划临床医学领域项目(19411951200)
苏州市科技发展计划项目(SS202072).
关键词
脑动脉瘤
自动检测
深度学习
模糊标签
双分支注意力
cerebral aneurysm
automatic detection
deep learning
fuzzy label
dual branch channel attention mechanism
