Automatic Extraction Of Msp In Brain MRI Based On 2-Channel And Self-Supervised Network

The approximate symmetry of brain changes with pathological changes,which may lead to the changes of brain structure or tissue structure.Due to the needs of practical application,the extraction of mid-sagittal plane(MSP)of brain image has an important influence on image registration,brain segmentation,pathological detection and medical image classification.Therefore,symmetry analysis of brain image is considered to be a promising and necessary technique.However,the symmetry detection of brain MRI image is still challenging,and after decades of research is still an important task.At present,there are two main types of MSP detection methods,the first is multi-target location based on symmetry,the other is multi-target location based on image recognition.However,most of the existing brain symmetry detection methods rely on local image feature detection and manual feature matching.Due to the complexity of feature selection of brain structure and the complexity of various noise products in brain MR images,the accuracy of traditional feature-based methods still can not meet the requirements of practical application.To solve these problems,this paper studies the automatic extraction framework of Mid-sagittal plane based on deep learning technology.Because the detection method based on symmetry relies on the characteristics of hand-made images and costs a lot of calculation,this paper proposes a two-dimensional brain MR image symmetry detection method based on 2-channel convolution neural network combined with deep learning.This is the first deep learning framework to extract fracture line by learning symmetry representation of human brain.Different from the traditional feature-based symmetry detection method,this method does not need any feature detection and specific matching operations.Using Poisson sampling to extract several pairs of brain image blocks randomly and evenly,the SPP structure in CNN framework can deal with the problem that these image blocks may have different resolutions.Using two channels of our two channel convolution neural network,including high-resolution and low-resolution channels,to compare the similarity of brain plaques based on the central axis and surrounding structures.Finally,the optimal axis of symmetry is determined according to the scoring and ranking scheme.Then,for the problem that the traditional algorithm based on imagerecognition is difficult to deal with various complex noises and heavy computation,this paper proposes an algorithm based on self-supervised learning network feature points and descriptor matching to extract the brain’s Mid-sagittal plane,which is also suitable for a large number of multi view geometry problems in computer vision.Unlike the neural network based on image block patches,the full convolution model of the self-supervised network runs the image on full-scale and computes the pixel level interest point location and related descriptors jointly.With the self-supervised training method,repeatable,stable and dense feature points can be trained.By matching the feature points and descriptors of a single MR image,the middle sagittal plane of the brain is obtained by voting,and the running time is greatly reduced.In this paper,the experiments based on 2-channel neural network and self-supervised network feature points and descriptors matching method are compared.The experiments show that our method has higher accuracy than the traditional method,more accurate extraction of brain midsagittal plane,and shorter running time.