Research On Semantic Segmentation Algorithm Of Brain Tissue Based On MRI

Brain is the most important organ of human body,and some unhealthy lifestyles may lead to various brain diseases,which seriously threaten human health.In clinic,MRI is usually used to image the brain structure,and then doctors distinguish each brain tissue in MRI,so as to analyze brain diseases.However,it takes a long time to distinguish brain tissues manually,which requires high professional knowledge and poor reproducibility.Therefore,it is very important to realize automatic and accurate semantic segmentation of brain tissues.At present,there are two problems in 3D MRI semantic segmentation network based on deep learning method: 1.The ability of feature extraction of multi-scale brain tissue or lesion area is poor,and the ability of small target detection is weak;2.3D MRI medical data set has high labeling cost and less labeled data.To solve the above problems,this paper proposes new methods for brain tissue semantic segmentation.The specific contents are as follows:Firstly,a new 3D MRI multimodal neurodegenerative disease data set is constructed.The data set included 50 sets of MRI images from 50 participants,each containing 2 modes,covering 5 types of samples: normal,alzheimer’s disease,mild cognitive impairment,dementia and parkinson’s disease,and six types of segmentation targets: 1)frontal lobe,2)temporal lobe,3)parietal lobe,4)hippocampus,5)midbrain and 6)centrum semiovale.Secondly,a new MRI brain tissue semantic segmentation method based on supervised learning is proposed.Aiming at the fusion of multimodal data,a new multi branch parallel residual module,attention module and hierarchical pyramid fusion module are introduced to effectively solve the problems of multiobjective segmentation,multi-scale segmentation and fuzzy segmentation edge.Finally,a semantic segmentation method of MRI brain tissue based on self-supervised learning is proposed.A large number of unlabeled images are used for model pre-training,so that the feature extraction network has strong feature extraction ability.Then,the downstream task of semantic segmentation is trained on a small number of labeled data set to effectively solve the problem of semantic segmentation of small data set.For the proposed method,rich experiments are carried out on the constructed basic data sets and public data sets,and compared with the existing methods,better results are obtained,thus confirming the effectiveness of the method in this paper.