Current Issue Cover
多通道融合可分离卷积 神经网络的MR脑部图像分割


摘 要
目的 卷积神经网络的方法可以提取到图像的深层次信息特征,在脑部磁共振图像(Magnetic Resonance Imaging,MRI)图像分割领域展现出了优秀的性能。但大部分深度学习的方法都存在参数量大,边缘分割不准确的问题。为克服上述问题,本文提出一种多通道融合可分离卷积神经网络模型(Multi-channel fusion separable convolution Neural Networks,MFSCNN)分割脑图像。方法 首先,在训练集中增加待分割脑结构及其边缘像素点的权重,强制使网络学习如何分割脑结构边缘部分,从而提升整体脑结构分割的准确率。其次,引入残差单元,以避免梯度弥散,同时使用深度可分离卷积代替原始的卷积层,在不改变网络每个阶段特征通道数的情况下,减少了网络训练的参数数量,降低了训练成本,减少了网络的训练时间。最后,将不同阶段的特征信息合并在一起,进行通道混洗,得到同时包含深浅层次信息的增强信息特征,加入到网络中进行训练,每个阶段的输入特征信息更丰富,学习特征的速度更快,收敛速度更快,显著地提升网络的分割性能。结果 在IBSR数据集上的分割性能表明,MFSCNN方法的相对于普通CNN的方法要明显提高,且在边缘复杂的部分,分割效果更理想,Dice和IOU值分别提升了0.9%-6.6%,1.3%-9.7%。在边缘平滑的部分,MFSCNN方法比深层次网络ResCNN和DenseCNN分割效果要好,而且MFSCNN的参数量仅为ResCNN的50%,DenseCNN的28%,在提升分割性能的同时,也降低了运算复杂度,缩短了训练时间。同时,在IBSR,Hammer67n20,LPBA40三个数据集上的表现,MFSCNN的分割结果和现有的其他主流方法相比,性能都更出色。结论 本文提出的MFSCNN方法,加强了网络特征的信息量,提升了网络模型的训练速度。实验表明,对于不同数据集,都能获得比较精确的MR脑部图像分割结果。
Multi-channel fusion separable convolution Neural Networks for Brain MR Image Segmentation

guotongyu,wangbo,liuyue,weiying(Northeastern University)

Background Brain is the nerve center of the human body, which controls human thinking and emotion. It contains many complex anatomical structures, including the hippocampus, putamen, caudate nucleus and so on. Their structural changes are closely related to many brain diseases. Magnetic resonance imaging (MRI) can perform high resolution imaging of brain and nervous system and other soft tissues, which has the characteristics of non-invasive, painless and fast acquisition and has become an effective tool for diagnosis, tracking, treatment evaluation and monitoring of brain development. It is also an effective clinical method for brain structure analysis. Objective Convolution Neural Network shows excellent performance in the field of Brain Magnetic Resonance image segmentation owes to its ability to extract the deep information features of the image. However, majority deep learning methods have the problems of too many parameters and the result of edge segmentation is inaccurate. To overcome these problems, a Multi-channel fusion separable convolution Neural Networks (MFSCNN) is proposed in this paper. Method Firstly, in order to make the network acquire more features of brain structure and its edge in training, the weight of the brain structure and its edge pixels is increased in the training set, and the network is forced to learn how to segment the edge part of brain structure, so as to improve the accuracy of the whole brain structure segmentation. Secondly, the residual unit is introduced to allow the network to transfer the derivative back to the network by jumping connections between the layers of the residual network. While deepening the network, the gradient dispersion can be avoided, which makes up for the lack of information loss in information transmission. At the same time, the deep separable convolution is used to replace the original convolution layer and the depth is used to replace the width. Without changing the number of characteristic channels in each stage of the network, the number of network parameters is greatly reduced, the number of network training parameters is reduced, the training cost is reduced, and the training time of the network is reduced. Finally, the feature information of different stages is merged together, and the channel is shuffled to get the enhanced information features which contain both deep and shallow information, and then put them into the network for training, and the input feature information of each stage is richer, the learning feature is faster, the convergence is faster, so that the performance of brain image segmentation based on network is obviously improved. Content of main experiment and result Firstly, for IBSR data sets, the results of MFSCNN are compared with those of ordinary convolutional neural network model (CNN), neural network model with residual unit (ResCNN) and neural network model with local full connection (DenseCNN). The network structure used in this paper is divided into four stages, each stage is a specific unit. In training and testing, 75% of the samples were selected as training set and 25% as test set. In this paper, Dice value and IOU value are used to measure the accuracy of image segmentation. Dice value can measure the similarity between the segmentation results and gold standard results. IOU value reflects the coincidence degree between the segmentation results and gold standard results. The experimental results on the IBSR dataset show that the results of MFSCNN are significantly higher than results of CNN, and in the complex part of the edge, the performance of segmentation is improved obviously. Dice and IOU increased by 0.9% - 6.6% and 1.3% - 9.7% respectively. In the edge smoothing part, MFSCNN is better than deep network ResCNN and DenseCNN in segmentation effect. Moreover, the parameters of MFSCNN are only 50% of ResCNN and 28% of DenseCNN, which not only improves the segmentation performance, but also reduces the computational complexity and training time. Comparisons with reviewed researches. In the performance on the IBSR, Hammer67n20, and LPBA40, the segmentation results of the method of MFSCNN are better than other existing methods. Among them, the MFSCNN method is more prominent in the segmentation of hippocampus. Compared with the commonly used segmentation software FIRST and FreeSurfer, the average Dice value of putamen and caudate nucleus is increased by 3.4% and 8%, respectively. At the same time, for the popular methods Brainsegnet and MSCNN+LC increased by 1.6%-4.4% and 2.6%-2.7%, respectively. Conclusion The MFSCNN method proposed in this paper can form a more friendly initialization training set for brain structure segmentation by increasing the weight of the interested brain structure and its edge pixels in the training set. When training the network, the deep separable convolution structure is used instead of the original convolution layer, which reduces the amount of network training parameters. At the same time, the feature maps of each stage are merged together and the channels are shuffled to get the enhanced information features that contain both deep and shallow information, which improves the accuracy of network model segmentation. MFSCNN not only solves the problem of inaccurate segmentation of complex edges of brain structure by traditional CNN, but also improves the inaccurate segmentation of lateral edges of brain structure by ResCNN and DenseCNN. In addition, experiments show that for different data sets, more accurate segmentation results of MR brain images can be obtained. Meanings: The regional contrast of MR image is low, and the gray value of each structure is very similar. Therefore, the fusion information can be extracted directly from MR image by MFSCNN method proposed in this paper, and further applied to other MR image segmentation. Although MFSCNN achieves good results for deep brain structure segmentation, the accuracy of segmentation for the discontinuous part of the brain structure still needs to be improved, mainly because of the complex and discontinuous types of pixels on the edges of these parts. Therefore, how to extract features that can segment complex edge contours using deep convolution network is also a problem that needs to be studied in the future.