Brain(functional) MR Image Segmentation And Structured Sparse Brain Functional Network Modeling

The development of brain imaging,particularly in the development of magnetic resonance imaging(MRI)technology,enables brain scientists to study the structure and functional activities of the brain in vivo directly and noninvably.Magnetic resonance imaging technology has become an important tool for human exploring the mysteries of the brain and studing neural pathology.It can provide vast amounts of medical data for the clinical diagnosis and pathological analysis of brain diseases,especially neuropsychiatric diseases and has become an important tool for research.With the maturation and development of medical imaging technology,brain lesion’s and tissues’segmentation of neuroimaging images for the diagnosis of brain disease,and brain functional network’s construction for the neuropsychiatric disease prediction and diagnosis,have become the new focus.Therefore,designing a robust and automatic image segmentation method for region of intreast,such as lesion,can dissipate the radiologists and medical doctors from the complicated work of manual processing of massive data and provide them lots of useful information extracted from the data.Study of brain functional network has shown great potential in understanding brain functions and identifying biomarkers for neurological and psychiatric disorders.Accurate construction of brain functional network from functional magnetic resonance images is an essential step prior to the subsequent statistical analysis or disease classification.Therefore,how to model functional connectivity network plays an essential role for accurate neuropsychiatric disease prediction and diagnosis.In this paper,we study the automatic segmentation of multiple sclerotic lesion as well as brain tissues by analyzing the multi-modality MRI data,and explore the structured sparse brain functional network to predict the neuropsychiatric disease based on brain functional magnetic resonance image.The main contributions of this thesis are as follows:(])By anlysising the intensity information feature of multiple sclerotic lesion in multi-modality brain MRI data(T1-weighted,T2-weighted,PD-weighted and FLAIR images),we propose to combine the supervised and unsupervised methods together to achieve the automatic segmentation of lesion.Firstly,an initial result that contains lesion is segmented by a supervised nonlinear voxel-wise classifier that trained based on intensity features extracted from multi-modality MRI sequences,and tissues’probabilistic prior from T1-weighted image.Secondly,with the prior that lesions almost exist in white matter regions and around the ventricle,some false-positive labels that located outside the white matter regions is discarded.Lastly,to further segment precise lesions boundary and detect missing lesions,we modify the region-based level set evolution,an unsupervised segmentation method,for finer and more accurate segmentation.The experimental results show that our proposal is robust and can be treated as an effective segmentation tool for multiple sclerotic lesion.(2)We present an automatic learning-based algorithm for segmentation of 3T brain MR images by learning segmentation information obtained from their corresponding 7T MR images.By iteratively training random forest classifiers based on the image appearance features and the context features of progressively updated tissue probability maps,a sequence of classifiers are trained with an auto-context strategy.With significantly higher intensity contrast and more anatomical details,7T MR images can provide more precise brain tissue information for classifier training.The experimental results on multiple real dataset have confirmed that the learned classifiers can progressively refine the tissue probability maps for achieving final robust tissue segmentation.(3)For brain functional network construction,we propose a connectivity-weighted sparse brain network model,which can make full use of the original BOLD singnal’s pairwise correlation by treating it as the measurement of connectivity strength during the sparse brain network modeling.To further make the connectivity strength-weighted penalty consistent across all links that have similar functional connectivity strength,we propose a group constraint on the similar links,for allowing them to share the same penalty during the whole brain network construction.In this way,we can model the whole brain network jointly,instead of separately modeling each ROI.By integrating connectivity strength,group structure,and sparsity in a unified framework,we can construct a more biologically meaningful brain network to obtain superior classification results in mild cognitive impairment identification.(4)We propose a graph regularized weighted sprase model to consider the potential structure during the brain network construction.Specifically,we introduce a graph laplacien regularization to constraint two similar signal data in the original data space should also maintain similar characteristics after projection.Then,we adopt our previous innovation,consnectivity-based weight penalty,for sparsity constraint and thus we get the final construction model that combines correlation-based analysis,sparsity constraint and graph regularizated constraint in a unified framework.For the feature selection part,we propose a sparse representation method to select the discriminant features of all the sets of brain network connectivities.This feature selection method can take into account the interaction between the features and then extract a subset of discriminant features.In this way,the final classification results can be improved to a certain extent.