Evaluation of a quantitative metric, Volumetric Statistical Amyloid Burden (VSAB), for Florbetapir PET, to classify amyloid positive and negative subjects using a cutoff derived from an independent dataset

511 Objectives Previously, it was shown on a large dataset of 183 subjects that VSAB, the volume of gray matter that exceeds a z-score threshold when compared to a database of young healthy controls, provided good accuracy for distinguishing between amyloid+ and amyloid- patients (Piper et al. EANM 2014). Our goal in the current work is to use the cutoff that was derived from this analysis and evaluate the accuracy on a large independent dataset from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Methods Florbetapir PET scans for 352 subjects from the ADNI were selected that had visual expert reads establishing the classification of amyloid+ or amyloid-. MIMneuro was used without intervention to deformably register each scan to a common template space. Z-scores were computed for every voxel after comparison to 74 young healthy controls within a probabilistic gray matter mask (GMM). The volume cutoff of voxels greater than the threshold z-score of 7 as determined from the previous study was applied. Kappa values and accuracy relative to the expert visual reads were calculated. Results The VSAB value used, previously derived from an independent dataset, for a z-score threshold of 7 was 1.46%. Out of 352 total subjects tested using a z-score threshold of 7, 139 of 150 visually positive subjects were classified correctly as positive (92.7%) and 196 of 202 visually negative subjects were classified correctly as negative (97.0%), for an overall accuracy of 95.2% and kappa=0.90. This compares favorably to the results of the training set from the previous study with an accuracy of 97% and the agreement was 0.94. Conclusions VSAB provided excellent agreement with expert visual assessment on a large dataset using a cutoff determined from an independent training set, suggesting this metric may provide a robust and accurate method for the quantitative assessment of amyloid images.