Hippocampus MRI Segmentation

Brain disorder is one of the main diseases that threaten the health of human body.Thus,it is necessary to study on early diagnosis of brain disorder which can hold back and delay the process of brain disorder,a small progress can bring a great help to the person or the society.In clinical,hippocampus volume plays an important role at the dysfunction and neurodegeneration of many brain disorders.Many studies support that an altered hippocampus volume and it's connectivity represent specific endophenotype at Alzheimer's disease(AD),schizophrenia and mild cognition impairment and etc.Hippocampus is very important in the study of the human brain's memory.Hippocampal morphology alteration and volume shrinkage are regarded as the key index to memory damage.As a biomarker of AD early diagnosis,the study of hippocampal morphology alteration on brain magnetic resonance image(MRI)has been the research hotspot in image.Hippocampus segmentation on brain MRI is a significant step to study the hippocampus volume shrinkage,a key issue and basis on hippocampal volume measurement and3 dimension reconstruction.And it has important application value at clinical diagnosis,therapeutics,evaluation and computer-aided diagnosis of brain disorders such as AD,schizophrenia and amnestic syndrome.At the same time,it has important value on the clinical diagnosis and periodization of AD,and it is also as efficiency index of an imaging bio-marker,therapeutic evaluation and prognosis of Alzheimer's diagnosis.Therefore,As a non-invasive imaging tequnique,it has important significance to carry out the research of the hippocampus segmentation on brain MRI.The current situation and development trend of hippocampus segmentation method are mostly image-based and model-based,and active contour model has been one of the wildly used methods owing to the advantages of aptness to process topology change,extend to high dimension and integrate the prior shape.However,due to the weak edge,the low tissue contrast,partial volume effect of Brain MRI,also due to the complex topological structure and the smaller hippocampus volume compared to the whole brain,it is very difficult to segment the hippocampus accurately and fast.We need to seek a method to segment hippocampus successfully.In our paper,we reviewed and analyzed the current methods of hippocampus segmentation.To avoid the disadvantage of these methods,such as the complex processing,the heavy computation burden,through synthetical consideration with the anatomy structure of the brain limbic system and the characteristics of brain MRI,we proposed a lattice Boltzmann(LB)method to segment hippocampus of brain MRI by constructing a local statistical region information and a prior shape information as an exterior force into LB diffusion equation.We select AD and AD's MRI of brain diseases from ADNI(Alzheimer's disease neuroimaging initiative)database as a main study object to segment the hippocampus.By choosing human control data(36)as the training set,hippocampus MRI of AD(41)and mild cognition impairment(38)are segmented and evaluated with the ground truth using Dice coefficient and computation time and iteration numbers.The experiments results show that:(1)the local statistical LB algorithm achieved better segmentation performance(Dice Coefficient(DC)= 96.8%)during segmenting brain real MRI with low tissue contrast and restrained the noises compared to other level set algorithms(DC=92.8%),furthermore,the novel algorithm improved the segmentation speed2-10 times than the art-of-the-state level set models;(2)the proposed algorithm based on shape moments had the advantages of simplity and efficiency to the description,extraction and incorporation to LB segmentation method,thus our algorithm is simple and the average time consuming is little(a second or so),which decreased by 5-140 times compared with the classical registration algorithms;(3)the proposed method based on region and shape information achieved the hippocampus segmentation on MRI.The average Dice coefficients can be reached 0.8,which is much higher than the existing LB method;But the iteration numbers are 2-6 times less than the art-of-the-state of level set method on hippocampus segmentation,and the computer time decreased by 5-16 times compared with the level set methods.The efficiency of the proposed method is improved much better.The main work and innovation of this paper are as follows:(1)we assumed a primary LB segmentation model applied to hippocampus segmentation on brain MRI,including the key elements: the diffusion coefficient and the exterior force.We also explained the physical meaning and the construction techniques of these key elements.(2)We proposed a novel local statistical LB algorithm(LSLBA)assembled by local statistical region information,which can increase the between-class variances of the foreground and background by reducing intra-class variations and can achieve a better anti-noise performance by smoothing the noise of neighborhood pixels.(3)We presented a segmentation model combining a prior shape described by geometric moments,which have invariance properties of translation,scaling and rotation and the training sets can be registered to extract the 'average'shape.Then we used the normalization to align the source object to the target object when incorporating the shape prior to the evolution model,here Euclidean distance is as a similarity measurement.By this way the registration progress is simplified and the efficiency.(4)Considering the characteristic of hippocampus MRI,we proposed a new LB model combined with a local statistic region information and shape prior information as the exterior force.Our method realized the hippocampus segmentation and it is improved at the segmentation efficient and accuracy compared with other methods.But the precise shape is one of the key factors on the construction of the prior information,it still need to be further improved.