基于三维卷积神经网络的颈部TOF MRA图像的血管自动分割

Automatic vessel segmentation of neck TOF MRA image based on 3D convolutional neural network

目的提出一种基于三维卷积神经网络(convolutional neural network,CNN)的分割模型,通过设计网络结构和调整参数,实现对颈部时间飞跃法磁共振血管造影(time of flight magnetic resonance angiography,TOF MRA)图像的自动精准血管分割。方法三维颈部TOF MRA图像数据来源于一项无症状老年人心脑血管病发病风险研究中的166例受试者。首先由有血管影像临床经验的放射诊断医生运用Mimics软件对颈部动脉血管进行手动标注,再按照8∶1∶1的比例将数据随机分为训练集(132例)、验证集(17例)和测试集(17例)。针对TOF MRA图像稀疏性和高对比度的特点,采用3D CNN的优化模型,通过在U-Net网络的编码和解码路径中多层级的加入专门的模块和不同分辨率的原始图像,更好地学习和利用到图像的独有特征。为研究输入图像的大小对网络性能的影响,从原始图像中分别裁剪出3种不同尺寸大小的三维切块来进行模型训练。在十折交叉验证下,采用Dice系数的平均值和标准差,以及灵敏度和特异度对模型的分割性能进行评价。采用单因素方差分析对比3种切块尺寸下的测试实验。结果所提出的新模型取得了最高的分割Dice值(0.9320)、灵敏度(0.9186)和特异度(0.9996),以及最小的Dice值标准差(0.0051)。ANOVA显著性水平为P<0.01,表明了不同切块尺寸下的模型分割结果有显著性差异,尺寸增大,模型分割结果更好。结论所提出的基于3D CNN的优化模型在TOF MRA图像血管自动精准分割上优于已有方法。此外,增加模型输入图像的尺寸大小有助于提高分割性能。

Objective We propose an automatic segmentation model based on 3D convolutional neural network to segment neck vessel from the time-of-flight magnetic resonance angiography(TOF MRA)data.Methods A total of 166 neck TOF MRA images were selected from the Cardiovascular Risk of Old Population(CROP)study.Vessel segmentation labels were manually delineated by vascular imaging professionals in 3D TOF images(Materialise Mimics,Mimics Medical 17.0),and were subsequently examined by experienced imaging experts.Randomly,132 images were assigned into training dataset,17 images were assigned into validating dataset,17 images were assigned into test dataset according to a ratio of 8∶1∶1.Based on the sparsity and high contrast of TOF MRA image,a 3D CNN model suitable for TOF MRA data were proposed by adding special modules and images to the encoding and decoding path of U-Net network.The unique features of the image could be better learned and utilized.In addition,in order to study the influence of the size of the input on the network performance,three different sizes of 3D patch randomly cropped from the original image were selected for model training and predicting.The average Dice coefficients,sensitivity,specificity and standard deviation of the Dice coefficients under 10-fold cross-validation were used to evaluate the segmentation performance of the model.In this comparative experiments,One-way ANOVA was carried out for the experiments of the three kinds of patch size.Results Proposed 3D CNN model obtained the highest segmentation Dice coefficient value(0.9320),sensitivity(0.9186)and specificity(0.9996),and the smallest standard deviation(0.0051).In the results of One-way ANOVA,the P<0.01,indicated that the results of different patch sizes had significant differences.And as the size increased,the segmentation result of the model was better.Conclusions The experimental results showed that the proposed model could automatically segment the neck vascular from TOF-MRA volumes and outperformed the state of the art.Besides,increasing the size of the input image could improve the segmentation performance.