基于图卷积网络的脑胶质瘤核磁共振图像分割

Magnetic Resonance Image Segmentation of Gliomas Based on Graph Convolution Network

近年来基于卷积神经网络(CNN)的图像分割应用已十分广泛,在特征提取的部分取得了很大进展.然而随着卷积层数越来越深,感受野不断增大,使模型丢失局部特征信息进而影响模型性能.使用图卷积网络(GCN)处理图数据结构的信息,能够在保留局部特征同时不随层数的加深而丢失局部信息.本文主要研究将基于CNN结构的对称全卷积网络(U-Net)特征提取与基于GCN的图像分割结合,提取全局与局部、浅层与深层的多尺度特征集应用于多模态脑胶质瘤核核磁共振(MR)序列图像分割,可分为两个阶段:第1阶段利用U-Net对多模态脑核磁共振胶质瘤MR序列图像进行特征提取,通过多个池化层实现多尺度特征提取及上采样进行特征融合,其中底层输出较低级别特征,高层输出更加抽象的高级特征;第2阶段通过膨胀邻域及稀疏化处理将U-Net获得的特征图数据转化为GCN所需的图结构数据,将图像分割问题转化为图节点分类问题,最后通过余弦相似度量对图结构数据进行分类.在BraTS 2018公开数据库上的实验结果取得分割准确度0.996、灵敏度0.892的效果.相比其他深度学习模型,本方法通过多尺度特征融合,利用GCN建立高低级别特征的拓扑连接,确保局部信息不丢失以取得较好的分割效果,能够胜任临床脑胶质瘤核磁共振图像的分析需求,进而有效提高脑胶质瘤诊断精度.

In recent years,image segmentation applications based on convolutional neural networks(CNNs)have been quite extensive,and great progress has been made in feature extraction.However,with convolutional layers increasingly deep,the receptive field is continually enlarged,which makes the model lose local feature information and affects model performance.Using graph convolution network(GCN)to process information on graph data structures preserves local features without losing local information as the layers deepen.This study focuses on combining U-Net(a kind of symmetric full convolutional networks)feature extraction based on CNN structure with GCN-based image segmentation to extract global and local,shallow,and deep multi-scale feature sets for multimodal glioma MR sequence image segmentation.The process can be divided into two stages.Firstly,U-Net is used to extract features from brain multimodal glioma MR sequence images,and multiple pooling layers are used to realize multi-scale feature extraction and upsampling for feature fusion,in which the bottom layer outputs lower-level features,and the top layer outputs more abstract high-level features.Secondly,the feature map data obtained by U-Net is converted into the graph structure data required by GCN by dilating neighborhood and sparsification,and the image segmentation problem is converted into the graph node classification problem.Lastly,the graph structure data is classified by cosine similarity.Experimental results achieved segmentation accuracy of 0.996 and sensitivity of 0.892 on the BraTS 2018 public database.Compared with other deep learning models,this method,by fusing multi-scale features and using GCN to establish topological connections between high and low level features,ensures that local information is not lost to achieve better segmentation results,which meets the needs of analyzing clinical glioma MR images,and then effectively improves the diagnostic accuracy of gliomas.