脑血管数字减影血管造影高分辨率分割网络设计
Design of a high-resolution segmentation network for digital subtraction angiography of cerebral vessels
崔颖 1付瑞 1朱佳 1高山 1陈立伟 1张广2
作者信息
- 1. 哈尔滨工程大学 信息与通信工程学院,黑龙江 哈尔滨 150001
- 2. 哈尔滨医科大学附属第一医院 神经外科,黑龙江哈尔滨 150001
- 折叠
摘要
针对现存卷积神经网络对脑血管数字减影血管造影分割精度不高的问题,本文提出了一种基于U-Net的改进网络(IC-Net).通过融合使用Inception和CAM通道注意力模块,以多种感受域提取更丰富的血管特征信息,并对特征信息进行筛选.增加 7×7 卷积层,通过压缩特征层分辨率的方式减少训练过程中产生的数据量.本文所提模型与U-Net、R2U-Net、Attention U-Net相比,IOU、Accuracy、F1-Score和ROC曲线下面积 4 项指标平均提升了1.82%、0.014%、1.19%和 0.73%.结果验证了IC-Net模型明显提升了脑血管数字减影血管造影虚弱血管和血管末端的检测能力,提升了分辨伪影噪声的能力,为医生识别脑血管中产生的病变提供有力参考.
Abstract
To solve the problem of low accuracy of existing convolutional neural networks for cerebral vascular DSA image segmentation,an improved network based on U-Net(IC-Net)is proposed.By fusing the use of inception and channel attention modules,rich vascular feature information is extracted using multiple sensory domains and feature information is filtered.A new 7×7 convolutional layer is added to reduce the amount of data generated dur-ing training by compressing the feature layer resolution.Compared with the U-Net and common U-Net improved models,the improved model's intersection over union,accuracy,F1-score,and area under the curve increase by 1.82%,0.014%,1.19%,and 0.73%on average,respectively.The results verify that the IC-Net model remark-ably improves the model's capabilities to detect weak vessels and vessel ends in cerebrovascular digital subtraction angiography images and distinguish artifactual noise.The model provides a strong reference for physicians to identi-fy lesions within cerebrovascular vessels.
关键词
图像分割/特征提取/脑血管/数字减影血管造影/U-Net/Inception模块/通道注意力/降维处理Key words
image segmentation/feature extraction/cerebrovascular/digital subtraction angiography/U-Net/in-ception module/channel attention/dimension reduction treatment引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据