Classification Of Mild Cognitive Impairment Based On Partial Volume Correction

Mild cognitive impairment(MCI)is a clinical condition between normal aging and Alzheimer's disease(AD).According to the development of disease,MCI can be divided into progressive MCI(pMCI)and stable MCI(stable MCI).Proper classification of MCI can help identify high-risk MCI patients in early clinical diagnosis.In recent years,positron emission tomography(PET)has become an important tool for studying AD.Due to the limitation of imaging theory and detection technology,however,the signal intensity of different pixels or regions will spread to each other in a PET image,known as the partial volume effect(PVE).PVE reduces the spatial resolution and degrades image quality,which leads to errors in the quantitative analysis and affects the diagnosis of diseases.Therefore,it is worth exploring to perform partial volume correction(PVC).In this study,298 normal subjects,100 AD patients and 214 MCI patients were selected from the ADNI database.Multi-modality data were acquired for each subject,including AV-45-PET,FDG-PET,and MRI images.And each MCI patient had a follow-up visit two years later after the first visit.Followed by image preprocessing,four PVC algorithms were performed on FDG-PET and AV-45-PET images: Van-Cittert deconvolution,Richardson-Lucy deconvolution,Muller-Gartner method and Geometric Transfer Matrix.The voxel features were obtained through region of interest and the cerebellum was selected as a reference region to normalize the voxels.Then the principal component analysis was used to reduce the dimension of data.Three machine learning classifiers were applied with kfold cross-validation,including support vector machine(SVM),decision tree(DT)and K nearest neighbor(KNN).The classification performance of MCI was evaluated by the receiver operating characteristic(ROC)curve.The higher sensitivity,specificity,accuracy and area under the curve(AUC),the better classification performance is.Results showed that the PVC algorithms had certain correction effects compared with the original PET images.The classification results of different modalities,different classifiers and different PVC algorithms showed that PVC could improve the classification performance of MCI under some circumstances.When achieving the highest prediction accuracy,AV-45-PET performed better than FDG-PET and could reflect the pathological features of AD more intuitively.According to the average evaluation metrics of different classifiers,the classification performance of SVM and KNN were relatively good,while DT gave poor results,which may not be suitable for classification of MCI.Therefore,appropriate PVC algorithm and classifier model can improve the classification performance of MCI to a certain extent.