Validation of FDG-PET datasets of normal controls for the extraction of SPM-based brain metabolism maps

Purpose An appropriate healthy control dataset is mandatory to achieve good performance in voxel-wise analyses. We aimed at evaluating [18F]FDG PET brain datasets of healthy controls (HC), based on publicly available data, for the extraction of voxel-based brain metabolism maps at the single-subject level. Methods Selection of HC images was based on visual rating, after Cook's distance and jack-knife analyses, to exclude artefacts and/or outliers. The performance of these HC datasets (ADNI-HC and AIMN-HC) to extract hypometabolism patterns in single patients was tested in comparison with the standard reference HC dataset (HSR-HC) by means of Dice score analysis. We evaluated the performance and comparability of the different HC datasets in the assessment of single-subject SPM-based hypometabolism in three independent cohorts of patients, namely, ADD, bvFTD and DLB. Results Two-step Cook's distance analysis and the subsequent jack-knife analysis resulted in the selection of n = 125 subjects from the AIMN-HC dataset and n = 75 subjects from the ADNI-HC dataset. The average concordance between SPM hypometabolism t-maps in the three patient cohorts, as obtained with the new datasets and compared to the HSR-HC standard reference dataset, was 0.87 for the AIMN-HC dataset and 0.83 for the ADNI-HC dataset. Pattern expression analysis revealed high overall accuracy (> 80%) of the SPM t-map classification according to different statistical thresholds and sample sizes. Conclusions The applied procedures ensure validity of these HC datasets for the single-subject estimation of brain metabolism using voxel-wise comparisons. These well-selected HC datasets are ready-to-use in research and clinical settings.

Automated summary

A well-selected healthy control dataset is essential for accurate voxel-wise analyses of brain metabolism. Different datasets show varying performance in assessing hypometabolism patterns in different patient cohorts, but overall accuracy is high.