Analysis Of Brain Structure And Risk Prediction Of Conversion In Patients With Mild Cognitive Impairment

Mild Cognitive Impairment?MCI?is a pre-dementia phase of Alzheimer's Disease?AD?,and its early diagnosis and conversion prediction are very important.In this study,brain structural MR images were collected from three groups including the healthy elders as normal comtrol?NC?,stable MCI?sMCI?and converted MCI?cMCI?,the follow-up data were acquired at three time points of baseline?first scan?,1 year and 2 years.First,the MR images were preprocessed and brain structural features were calculated by the FreeSurfer software.Then the differences of the cortical surface features?surface area,cortical thickness,gray matter volume,and mean curvature?,and the intensities of ROI-based hippocampal subregion voxels among three groups were extracted using the T-test method,and analysis of variance?ANOVA?was adopted to reveal the changes of cortical features for each group within 2 years.The experimental results showed that there were progressive changes in the brain structure of the patient from NC to sMCI,and from sMCI to cMCI.The changes occurred from the medial temporal lobe,part of the frontal lobe and the parietal lobe to the frontal lobe,the parietal lobe and the occipital lobe,and hippocampal subregions had atrophies from subiculum,presubiculum,CA2₃ and CA4 to the entire hippocampus structure.With the increasing of age,the changing rate of brain structural features were followed as cMCI> sMCI> NC.Second,T-test,sparsity-constrained dimensionality reduction?SCDR?and recursive feature elimination?RFE?were explored for feature selection,one type of features and their combination were considered as the input of support vector machine?SVM?for training and testing,so as to classify three groups and obtain the features with strong discriminative power.In classification task using SVM classifier,the RFE achieved best performance,the cortical thickness and the gray matter volume had better discriminative ability than the surface area and mean curvature,and hippocampus voxels exhibited the similar capability in classification.Combination of different features,and combination of baseline features with longitudinal features were compared in classification.It demonstrated that combined features could improve the performance of classifier.Finally,in order to further improve the classification performance and evaluate conversion risk,the deep learning algorithm was adopted for multi-classification and MCI conversion time prediction,and the performance of two convolutional neural networks and the SVM classifier were compared and discussed.With regard to deep learing architecture,the CaffeNet outperformed the GoogleNet and reached high accuracies with 90.46%,85.12%,91.22% for baseline NC,sMCI and cMCI,and obtained 96.74% accuracy in MCI conversion prediction.Therefore,this study illustrated that there existed significant differences in brain structures between NC,sMCI and cMCI groups.Deep learning networks had prominent ability in early diagnosis and conversion assessment of MCI,and it had potentials in clinicial application.t of MCI,and it had potentials in clinical application.