P3‐185: THE EUROPEAN DTI STUDY IN DEMENTIA: A NOVEL FRAMEWORK TO TEST THE DIAGNOSTIC USE OF DTI IN ALZHEIMER'S DISEASE

Diffusion tensor imaging (DTI) is a potential early marker for microstructural white matter degeneration in Alzheimer's disease (AD), but its use in a multicenter setting has not yet been investigated. The European DTI study on Dementia (EDSD) permits to study the stability and diagnostic accuracy of multicentre DTI. The EDSD has conducted a multicenter clinical and physical phantom study across 16 different scanners, and collected DTI and structural MRI data from 168 subjects with MCI, 137 patients with clinically probable AD and 196 healthy controls across 10 different MRI scanners. Diagnostic accuracies were compared between fiber tracking along the posterior cingulate, voxel-based random-effects analysis, voxel-based meta-analysis, and machine learning classification using support vector machines (SVM) to discriminate patients with AD dementia from healthy controls. SVM classification was then applied to discriminate CSF-positive MCI subjects from CSF-negative MCI subjects and healthy controls. We found 50% higher variability between scanners for voxel-based FA measurements compared to conventional morphological MRI (Figure). We found significant reduction of FA and significant increase of MD in core areas of AD pathology using both univariate and SVM based analysis in AD dementia and CSF positive MCI subjects. The effects of group were spatially more restricted with random effects modelling of scanner effects compared to simple pooled analysis. Posterior cingulate tractography yielded lower diagnostic accuracy comparing AD dementia and healthy control subjects than the voxel-based univariate analyses. The level of accuracy was highest using SVM based classification (about 80%) for all DTI parameters when comparing AD dementia cases from controls. When applying SVM analysis to prodromal AD, we found an accuracy of 62% for FA and 77% for MD, for the discrimination of amyloid-β42 positive MCI subjects and controls. For the group separation between amyloid-β42 positive and negative MCI subjects, we obtained a mean accuracy of 67% for FA, and 64% for MD. DTI is more affected by between-scanner variability than conventional MRI. Non-linear multivariate machine learning approaches seem superior to univariate models to account for between scanner effects and to increase diagnostic accuracy in AD dementia and prodromal AD. Coefficients of variation of FA across scanners Voxel-based coefficients of variation (CV) from one single subject scanned on 16 different scanners