Research Of Computer Aided Diagnosis For Mild Cognitive Impairment Based On3D Reconstruction And Machine Learning

Asworldwideagingofthepopulationbecomesserious,thevictimsofAlzheimer’sdisease put heavy burden to the whole society. With the fast development of com-puter science, especially computer graphics, digital image processing, machine learn-ing, Magnetic Resonance Imaging, which is one of the major approaches of studyingand diagnosing Alzheimer’s disease, has shown more and more potential. All of theachievements make great positive impact to the diagnosis and prevention of AD.The increasing of the speed of GPU is signifcant by contrast to the speed of CPU.In the meantime, as more realistic rendering are expected, graphics programming tech-nology are becoming more fexible. MRI images that are taken with small and fxedtranslation can be constructed to vivid3D objects, which will facilitate the doctors toobserve the whole brain directly. We introduce the3D MRI reconstruction softwarebased on GPU accelarated Ray-Casting algorithm, and explain the displine of tha corealgorithm and details of implementation.What’s more, the whole process of cortex reconstruction with Freesurfer are alsointerprated. Freesurfer is very powerful software about cortex reconstruction, whichcan get accurate geometry representation of the surface of cortex from the MRI imageseries. Afterthat, thethicknessofdiferentregionofinterestinthe cortexcanbecalcu-lated easily. The main feature of Alzheimer’s disease and Mild Cognitive Impairmentare the shrinking of the cortex, so the pathological changes can be measured diferenceof geometry feature.Employing the machine learning methods to help the automated diagnosis of ADand MCI has been popular in recent years. For feature selection, we adopt thickness of diferent cortex region. As for the classifcation methods, we performed a seriesof experiments, which includes feature subset selection, principal component analysis,selection of classifcation algorithm, under the classic machine learning framework.When dealing with the selection of classifcation algorithms, at frst we try classic ma-chinelearningalgorithms, suchaslineardiscriminantanalysis, supportvectormachineetc at frst, etc. And then, we use mixture of gaussians to model the probability distri-butionofeachclass, andgetbetterperformancethantheclassicones. What’smore, wealso explore the relationship between the number of components in the mixture modeland the classifcation accuracy.