P2‐234: Automatic detection of Alzheimer's disease in multicenter DTI data

Alzheimer's disease (AD) patients show significant changes in white matter (WM) structural integrity. Diffusion tensor imaging (DTI) is a neuroimaging technique that allows in vivo assessment of WM fiber tract integrity and, thus, could support the diagnosis of AD as an additional biomarker. Current research focuses on machine learning (ML) methods to automatically detect AD specific structural WM changes. Therefore, the algorithms used must be robust and stable to work with data recorded across different scanners. Within the newly created framework of the European DTI study in Dementia (EDSD) we have collected data of more than 330 subjects from ten scanners located at nine sites. Objective: To assess the accuracy of ML classifiers for the detection of AD based on a large multicenter DTI data set using different approaches to reduce inter-site variability. After strict quality control we pooled the remaining 280 DTI and MRI scans derived from 137 patients with clinically probable AD and 143 healthy elderly controls. For classification we used fractional anisotropy (FA) maps and mean diffusivity (MD) maps and performed a tenfold cross validation. We selected discriminative voxels using the information gain criterion and classified the data with a Support Vector Machine. In a second step, we eliminated variance attributable to center and other covariates including age, education, gender, using principal component analysis (PCA) before repeating the classification procedure. For FA and MD the feature selection identified areas in the medial temporal lobe and corpus callosum that had the strongest contribution to the group separation. We achieved an accuracy of 80% for FA and 83% for MD. For the tissue density maps we obtained 83% for WM and 89% for gray matter. The reduction of variance components arising from center, gender, age and education effects did not significantly change the classification results for FA and MD. Multicenter acquisition of DTI data in combination with multivariate ML approaches show promising results which can be compared to earlier monocenter DTI studies. Variance introduced by different scanners can be detected by PCA, but it seems not to affect the performance of the classifier.