Study On Brain Structure Segmentation Based On Fully Convolutional Neural Network

Brain diseases have threatened humans for many years.Alzheimer's disease(AD)is the one that is difficult to be diagnosed and cured in elderly brain diseases.It can cause problems such as memory loss and emotional disorders in brain.AD currently exhibits varying degrees of brain structural atrophy and expansion at different stages.Therefore,rapid and accurate brain structure segmentation techniques can help clinically find bio-imaging targets and locate lesions.At present,the brain structure segmentation method is mainly based on the atlas registration method.The accuracy of the rigid registration segmentation method is low,and the non-rigid method is slow.Both are affected by the differences between the standard map and the target image,which cannot be well applied in big dataset.In recent years,deep learning technology has made breakthroughs in disease identification,lesion localization and segmentation.The segmentation and registration method based on fully convolutional neural network provides a new solution to the field.In this paper we constructed a standard data sample of brain structure maps based on the Alzheimer's Disease Neuroimaging Initiative based on the brain map proposed by Johns Hopkins University.The standard data sample is the basis of the experiment and the key to the evaluation of the segmentation method.This study,with the help of clinicians,used multi-map technology and manual combination to screen the data and established the standard sample dataset.In this paper,the semantic segmentation model of brain structure is constructed based on the fully convolutional neural network.In this paper we analyzed and contrasted the current nerual network,FCN and SegNet,which are popular in computer vision and have been proven to gain a good segmentation effect in the medical field.Among them,the segmentation effect of SegNet is better than that of FCN.It could obtain the DC value of 0.885 on average in level 1 and 0.848 in level3,which is higher than traditional brain structure segmentation method based on atlas-based segmentation method.The continuity of the image after 3D reconstruction was poor.In order to solve this problem,a specific network structure SUF-net was proposed.And this network improves the image continuity and the segmentation accuracy,which could reach 0.902 in level 1 and 0.861 in level 3.In this paper,the proposed segmentation network was used to assist with disease diagnosis,and an early diagnosis model of AD was established.According to the prior knowledge of clinical and scientific research,the degree of atrophy of AD patients in different stages of brain varies.The stage of AD is realized byidentifying the characteristics of the designated brain regions.For different experimental groups,we conducted three comparative experiments to verify the effectiveness of the proposed model and achieved early diagnosis of the disease with an accuracy rate of 98%.In summary,brain structure segmentation combined with the early diagnosis model,constructing an analysis and diagnosis tool for AD,achieving an end-to-end model from reading and processing to diagnosis.