Brain MRI Segmentation Based On Fully Convolutional Networks

Modern medical imaging technology has become the most important means of auxiliary diagnosis of clinical brain diseases.Magnetic resonance imaging(MRI)can capture high contrast and high resolution soft tissue images,which is an important way for structural brain analysis.For example,in the diagnosis of Alzheimer’s disease,brain tumors and other brain diseases,magnetic resonance imaging of brain tissue needs to be taken,and brain MR image segmentation and corresponding measurement are needed.However,manual segmentation to do this can be very time-consuming and requires a high degree of expertise,so automatic brain tissue segmentation will greatly assist in medical diagnosis and development.In this paper,combined with the advantages of fully convolutional neural network in feature extraction,brain MR image segmentation is studied from three aspects:infant brain tissue segmentation,adult brain tissue segmentation and adult brain tumor segmentation.It has positive scientific research value.In order to improve the ability of segmentation algorithm to locate and classify infant brain MRI images,an infant brain MRI segmentation algorithm based on residual connection and dilated convolution pyramid is proposed in this paper.In this algorithm,the classical U-Net network is improved,and the residual connection is used instead of the traditional stacked structure to effectively improve the representation and discrimination ability of the model.Then a multi-branch dilated convolution pyramid block is proposed to solve the problem of spatial information loss in the process of downsampling,improve the network location ability,and effectively aggregate multi-scale information.The algorithm has won the first place in many indexes of Iseg2019 infant brain tissue segmentation challenge.In order to correctly classify different tissues of adult brain,such as gray matter,white matter and cerebrospinal fluid,and accurately locate the boundary of each tissue,an adult brain MR image segmentation algorithm based on spatial self-attention and deep feature reconstruction is proposed in this paper.The algorithm mainly promotes the exchange of information between different convolutional layers,uses the high-level features to make the correct classification,and uses the low-level features to locate accurately.The self-attention module is used to encode the long range information,and the deep feature reconstruction module is used to modify the feature.The network self-learning mechanism enables the high and low level features to complement and strengthen each other.In this paper,the algorithm achieved the sixth place in the MRBrainS 18 Challenge.In order to segment brain tumors accurately and reduce the computational cost caused by three-dimensional convolution,a brain tumor segmentation algorithm based on separable convolution and channel shuffle is proposed in this paper.In this algorithm,three-dimensional separable convolution is used to reduce the computational complexity of three-dimensional segmentation network,and the three-dimensional separable convolution of three branches is used to extract and fuse axial,sagittal and coronal information.Channel separation and channel shuffle are used to futher limit the amount of network computation and integrate the information of different branches.In the case of very little computation and the number of parameters,the algorithm has an outstanding performance on the BraTS-2018 training set.Finally,this paper summarizes the research work carried out in this topic,and looks forward to the future research direction.