A Study On Brain Magnetic Resonance Image Segmentation Combining With Tissue Characteristic

Magnetic resonance imaging(MRI)is widely used in research of brain science and clinical diagnosis of brain disease due to its non-invasive and high soft tissue contrast properties.Segmentation of brain MR image can assist doctor diagnose brain disease.Longitudinal relaxation time T1 is an inherent property of MRI tissues.The tissue characteristic T1 mapping not only reflects the physiological or pathophysiological characteristics of the imaging tissue,but also provides tissue features which are not found in original brain MR image.This research mainly studies the brain MR image segmentation method combining with tissue characteristic.The research contents are described as follows:(1)Based on SE-IR(spin-echo-inversion recovery)sequence and inversion time,T1 mapping of scanned position is calculated and fitted to extract tissue characteristic.First,the IR sequence is used to pre-pulse the SE sequence to obtain T1 weighted images with different inversion times.Then,the non-linear least square method is used to fit longitudinal relaxation time T1.Finally,the T1 value of the brain tissue is colorcoded to obtain a T1 mapping that reflects tissue characteristic.(2)For the problems of brain MR images with unclear boundaries and uneven gray levels,a brain image segmentation method which combines tissue characteristic and fuzzy support vector machine is proposed.First,the T1 mapping that reflects the characteristic of brain tissue is determined by sequence measurement,and image label is obtained after preprocessing.Then,a membership function based on the maximum fuzzy distance between classes is designed to determine the membership of each sample.The determination of the membership function comprehensively considers the spatial distance between samples of different classes,and reduces the degree of membership dependence between samples of same class.Finally,the fuzzy support vector machine is trained to segment three main brain tissues.Segmentation experiments show that the proposed method can effectively and accurately segment brain tissue.(3)For the problem that convolutional neural network is easy to cause loss of feature information during the up-sampling and down-sampling process,resulting in unsatisfactory segmentation results,a brain image segmentation method is proposed which combines the tissue characteristic and the deep feature aggregation network(DFA-Net).DFA-Net is equipped with three levels of feature representation layers(basic layer,middle layer,and aggregation layer).The basic layer is used to extract basic features of original brain images.The middle layer is used to fuse tissue characteristic and feature information of brain images,and the feature from basic layer and tissue characteristic T1 mapping are aggregated.The aggregation layer is used to deeply aggregate the features of the middle layer to make up for the feature loss of upsampling and down-sampling.In addition,the Feature Aggregation Module(FAM)can utilize the complementary information of deep features and shallow features to automatically aggregate and activate shallow and deep features.Brain tissue segmentation experiments show that the proposed method can perform effective feature aggregation and improve the accuracy of brain tissue segmentation.