基于3D-RA图像的颅内动脉瘤自动检测算法

Automatic detection algorithm for intracranial aneurysm based on 3D-RA image

颅内动脉瘤检出率低、破裂后致死率高,是一种严重威胁人类生命健康的高发性脑血管疾病。针对二维卷积神经网络在动脉瘤诊断中对先验知识利用不足问题,基于3D-RA序列图像成像特点,提出一种基于光流可变形卷积的颅内动脉瘤检测算法。采用稠密光流算法获取序列图像之间的光流信息作为先验知识,结合光流信息改进二维卷积计算过程,提出光流可变形卷积模型,从而建立序列图像间的像素级联系。此外,结合光流可变形卷积和标准卷积组成编码模块,实现图像重要特征提取。以北京天坛医院360例临床3D-AR颅内血管造影数据为样本集,测试结果表明:所提方法正确率为0.9787、精确率为0.9836、召回率为0.9747、F1分数为0.9791、AUC为0.9924、mAP为0.9822;与传统网络U-net、Attention U-net相比,该网络对颅内动脉瘤检测更准确;与原有可变形卷积模型相比,光流可变形卷积模型利用光流作为先验知识,提高了网络性能。

Intracranial aneurysm which is a kind of high incidence cerebrovascular disease seriously threatening human life and health has low detection rate and high mortality after rupture.Aiming at the problem of insufficient utilization of prior knowledge by 2D convolution neural network in aneurysm diagnosis,an algorithm for intracranial aneurysm detection using deformable convolution integrated with optical flow is proposed based on the imaging characteristics of 3D-RAsequence images.Dense optical flow algorithm is used to obtain the optical flow information between sequential images as prior knowledge.Then the obtained optical flow information is used to improve the 2D convolution calculation process,and an optical flow deformable convolution model is proposed to establish the pixel-level connection between sequential images.In addition,an encoding module is composed of optical flow deformable convolution and standard convolution to extract important features from images.The 3D-AR intracranial angiography data from 360 clinical cases in Beijing Tiantan Hospital are taken as sample set,and the test results showed that the accuracy,precision,recall value,F1 score,AUC and mAP of the proposed method are 0.9787,0.9836,0.9747,0.9791,0.9924 and 0.9822 respectively.The accuracy of the proposed network in detecting intracranial aneurysm is higher than that of traditional networks U-net and Attention U-net.Compared with traditional deformable convolution model,optical flow deformable convolution model uses optical flow as prior knowledge,which improves network performance.