The Study And Application Of Multilevel Intelligent Classification Algorithm On Brain Images

With the development of medical imaging technology,brain structural and functional analyses have become hot research topics in recent years.Manually feature marking method and image analyzing approach increase clinician's burden.Extracting features from brain images using machine-learning method to diagnose brain disease and detect neuroimaging biomarkers has become research trend.Automatic feature extraction method and intelligent classification method can not only improve the diagnostic accuracy but also reduce the rate of misdiagnosis.During the process of diagnosing brain diseases,extracting optimal features that make the classifier perform well is the primary work.Therefore,studying automatic feature extraction method,especially constructing multilevel and high-dimensional features that reflect brain structural information comprehensively,is hot and difficult topic in research of intelligent classification algorithm.At present,the common features used in brain image analyses are either volume-based or surface-based.These single-level features cannot convey synthetical brain structural information.In this paper,we propose a multilevel-features-based intelligent classification algorithm to assist brain disease diagnoses and analyses.The main contributions and innovations are as follows:(1)Through the research on automatic feature selection methods,we propose a multilevel-features-based intelligent classification algorithm on brain images,which is useful for detecting neuroimaging biomarkers.The multilevel features consist of low-level ROI features(gray matter volume,cortical thickness,etc.)and high-level brain network features(connectivity between ROIs).Filter-based and wrapper-based feature selection method is used to reduce feature dimension.For each feature type,single kernel matrix is constructed using radial basis function(RBF).These two single kernel matrixes are integrated into a multi-kernel support vector machine(SVM)classifier using appropriate weighting factor.The multilevel features convey comprehensive information of the brain structural alterations.This classification algorithm has strong ability of generalization that can be applied to similar problems.(2)We study cerebral alterations of type 2 diabetes mellitus(T2DM)to find the statistical significant neuroimaging biomarkers using automatic image processing methods.T2 DM is a common metabolic disorder that causes irreversible brain tissue impairments and leads to complications,such as cognitive dysfunction.Neuroimaging plays an important role in T2 DM brain morphometry detection at the early stage of the disease.Previous imaging morphometric studies of T2 DM have mainly focused on the measurement of either volumetric or cortical characteristics,which is impossible to differentiate the potential abnormality among the three characteristics.With the state-of-the-art methods for brain segmentation and cortical surface reconstruction,we are able to compute the characteristics of the whole brain accurately,including gray matter volume,cortical thickness,and cortical surface area.These three measurements express different neuroanatomical meanings,i.e.,gray matter volume,constituted of neuro cells,reflects the total number of neuro,cortical surface area is determined by the number of neuron columns that run perpendicular to the cortical surface,whereas the cortical thickness is influenced by the number of cells within a column.Integrating these three measurements together is meaningful for thoroughly analyzing T2 DM related brain morphometry.(3)We examine the brain morphometry associated with self-esteem in young adults using multilevel-features-based classification algorithm.In order to understand self-esteem better,a kind of complicated cognition psychology,more researches are need on self-cognition neurological mechanism bases.Most of current neuroimaging researches use morphometric analysis method or voxel-based method to study self-esteem related brain structure.However,these methods can only study specific brain regions that deal with self-cognition information.Although these studies can give preliminary interpretation of the corresponding relation between processing process of self-esteem information and the brain regions,the neuronal network activities of specific brain areas are not studied.In our method,the multilevel features achieve promising classification results: accuracy = 96.66%,specificity = 99.77%,and sensitivity = 95.67%.Brain regions that are sensitive to self-esteem have been examined in our study.Our method not only repairs the deficiencies of existing methods but also provides information of both isolated ROI and brain connectivity between ROIs,which helps understand the development and change pattern under different self-esteem levels.The application of our method is important for both clinical and scientific research.(4)At present,there are many kinds of software that can assist brain disease diagnoses and medical image analyses at home and abroad.Although these toolboxes can perform image processing automatically,automatic feature extraction from brain images cannot be accomplished well.In this paper,we develop a brain image processing and analyzing system(BrainLab),which is able to process brain images automatically and extract multilevel features.This software platform can help clinician to early diagnose brain diseases.