Research On The Segmentation Method Of Brain Image Hippocampus

The hippocampus is an essential sub-structure in the human brain.The hippocampus has vital links with Alzheimer's diseases,schizophrenia,and many neuropsychiatric diseases.These diseases can cause damage to the hippocampus structure,and the volume will shrink gradually.By accurately measuring the size of the hippocampus,clinicians can find the degree of atrophy.Further,clinicians can diagnose and treat related mental diseases.The accurate measurement of hippocampal volume depends on the precise segmentation result.Therefore,the precise segmentation for the hippocampus in the brain image is of great significance for the study of various diseases.Domestic and foreign scholars have carried out a lot of researches on the segmentation of hippocampus.However,the hippocampus is relatively small,irregular in shape and the boundary with surrounding tissues is relatively vague,it was difficult to segment the hippocampus using structural magnetic resonance images accurately.The current segmentation method still has the following deficiencies.The dice similarities obtained by the current mainstream segmentation method were 0.85-0.90,and the segmentation process of the multi-atlas method with high segmentation accuracy was too dependent on registration accuracy and the segmentation time was too long,around 20-50 minutes,which increased linearly with the number of registration times.This is very unfavorable for the study of big medical data.To solve the above deficiencies,this thesis aims to realize the rapid and accurate segmentation of hippocampus in brain images.The method based on fully convolutional neural network which has been successfully applied to the segmentation of hypothalamus,caudate nucleus,and putamen of the brain four substructure.To solve the problems existing in the current segmentation,this thesis planned to use this method to segment the hippocampus and studied the repeatability of hippocampal radiomic features under the same sample with multiple scans.Specifically,the following studies were performed:(1)In this thesis,the fully convolutional neural network(FCNN)was used to segment the hippocampus in the brain image.To avoid redundant convolution and size operations,the network structure only consists of convolutional layers.This thesis makes a practical application of the network and verifies that the network can be used for the segmentation of the hippocampus.This thesis also provides a model of the hippocampus segmentation.The input brain image,the manually labeled hippocampus image,and the corresponding interest area image are used for network training,and the input process can accept input images of any size.The performance of the proposed method has been evaluated and compared with one multi-atlases method and the FreeSurfer method.For the segmentation of hippocampus,the dice similarities obtained by the segmentation network conducted by a leave-one-out cross-validation and ten-fold cross-validation all were approximately 0.91 for 65 subjects in the internal dataset.When the network trained by the internal dataset was used to segment the 135 subjects in the public dataset,the dice similarities was approximately 0.79.In the segmentation time of each subject,the multi-atlas method is close to one hour,and the method used in this thesis has a significant reduction over the segmentation time compared to the multi-atlas method,it takes nearly 10 minutes on the CPU and only 1 minute on the GPU for the segmentation of single hippocampus.The experimental results demonstrated that the fully convolutional neural network in this thesis could obtain significant improvement over the multi-atlases method in the segmentation time and can obtain a reliable accuracy.(2)This thesis analyzes the repeatability of the radiomic features of the hippocampus.Radiomic is a quantitative method to extract medical features,and it can extend details that are invisible to human eyes.In the current studies,it was found that the radiomic features of the hippocampus can well describe the hippocampus and can effectively identify normal people and AD patients through these features,but whether these features are stable in multi-scanning images,this is still a question that is worth exploring further.To explore whether the radiomic features of the hippocampus in the brain image are stable and which features can be used as stable biomarkers.In this thesis,four multiple scans of brain image datasets were selected,and the repeatability analysis of the intensity feature,the shape feature,the texture feature,and the wavelet feature were carried out.The Intraclass Correlation Coefficient(ICC)and Overall Concordance Correlation Coefficient(OCCC)were well studied in the four datasets.The experimental results show that the texture features of the hippocampus are the most repeatable and conclude that this feature may be an extremely reliable biomarker in clinical research.(3)This thesis designs an automatic extraction system for the hippocampal radiomic features.To study the hippocampus more conveniently and obtain the radiomic features that better reflect the hippocampus,we designed an automatic radiomic feature extraction system based on the two previous research.Currently,the system has the following four functions: one is the conversion of raw DICOM format data to Nii format;Secondly,the registration of brain image individual space and standard space;Thirdly,the rapid automatic segmentation of hippocampus based on fully convolutional neural network;Fourthly,the extraction of hippocampal radiomic features.The software is easy to use and can quickly realize image format conversion,image registration,hippocampus segmentation and the extraction of radiomic features of the hippocampus by using the visual button.