Research On Magnetic Resonance Imaging Based Methods For Mild Cognitive Impairment To Alzheimer’s Disease Prediction

Alzheimer’s Disease(AD)is an incurable neurodegenerative brain disorder,and the most common dementia in the aged population.According to report,AD patients’number has increased 68%at last decade,will affect 115.4 million people in 2050.When patient was diagnosed as AD,the irreversible damages have been caused.Mild Cognitive Impairment(MCI)is an inevitable stage of AD and a crucial basis for AD diagnosis.Recent research shows that MCI patients further progress to AD at a rate of 10%to 15%per year.MCI patients who not convert to AD or further progress to other forms dementia will remain stable,and in some cases,MCI patient could return to normal condition.Although existing treatments cannot cure AD,there some medicines could delay the onset of AD symptoms.This fact suggested that prognosis of AD should get more attention than diagnosis,and precisely predict a MC1 patient will or not progressive to AD,has the great significance for the MCI patients and its family,which also helpful to the AD clinical trial research.In this paper,we focused on the Magnetic Resonance Imaging(MRI)Based MCI to AD conversion prediction methods.For seeking more effective biomarkers,we first sorted out the existing related work and analyzed its shortcomings.Aiming at these deficiencies,we proposed three biomarkers:1)the coupling biomarker,which adopt a coupled feature representation method to express and utilize the linear and nonlinear correlation between features,to reduce the error of the assumption which regard features as independent;2)the deep biomarker,which employed a deep learning method to overcome the lack of inadequate expression ability of manual features;3)3D deep biomarker,which utilize a 3D convolution neural network to learn the space features in MRI and that could overcome the space information loss caused by using a 2D convolution neural network to dealing a 3D volume image.The experimental results show that the coupled feature predictor achieved accuracy of 75.4%which 3.3%increased over original feature,and verifies the existence of correlation between ROI features,this correlation also relates to the conversion of MCI to AD.In the sliced deep predictors,we first proposed the predict scheme which training sub-classifiers on MRI slices dataset and construct a voter by these sub-classifiers.Through extracts the high task-related region in brain template,we finds that the decision regions are distributed in the hippocampus,amygdala,spinal gyrus,pallidus,temporal pole which were corroborated with clinical findings.In the result,sliced deep predictor achieved accuracy of 81.1%which 5.7%increased over coupled feature predictor.The experiment of the spatial deep predictor show that 3D-CNN can learn spatial features more effectively,achieved accuracy of 83.1%,and in 3D-CNN based methods,regard AD and NC(normal control)samples as auxiliary data could effectively improve the prediction accuracy of MCI to AD conversion.Finally,we also introduced the cognitive measure scores which were combined with the biomarkers as a new feature,we used the new feature to training a random forest classifier for a second prediction as final result,and achieved state-of-the-art result.