Clinical Validation of Deep Learning for Image Restoration of Ultra-Low-Count [¹⁸F]FDG PET for Dementia Diagnostics

Deep learning (DL) represents a promising technique for image restoration. We explored its ability to restore ultra-low-count [¹⁸F]FDG PET studies of the brain in subjects with dementia and in healthy subjects to allow for reduced scan durations or administered activities without compromising diagnostic performance. <b>Methods:</b> Various DL models using the content aware image restoration approach of CSBDeep toolbox (3D U-nets) were trained with subvolumes of 1,000 random subjects. On the basis of 10-min list-mode PET data after injection of 208 ± 10 MBq of [¹⁸F]FDG, we reconstructed reduced scan durations of 2 min, 1 min, 30 s, 20 s, and 10 s. The resulting models were applied to [¹⁸F]FDG PET scans of subjects with Alzheimer disease (<i>n</i> = 15), frontotemporal dementia (<i>n</i> = 14), and healthy controls (<i>n</i> = 13). We explored the effect of reduced scan times on individual regional measures in diagnostically relevant regions and on voxel-based group contrasts. Three independent readers rated all datasets with regard to assessability, diagnosis, and diagnostic confidence. <b>Results:</b> Individual mean regional [¹⁸F]FDG uptake remained largely unchanged. The SD strongly increased with shorter scan duration without application of DL (mean increase ≤ 48%), whereas it slightly decreased with DL (≥-7%). In group contrasts, the number of significant voxels strongly decreased with shorter scan time without DL (≥-41%), which was partially offset by DL (≥-27%). On visual reads, the fraction of assessable images steeply fell to only 4% (10-s scan) for scan durations below 2 min without DL, whereas every single image restored with DL was assessable. The diagnostic confidence continuously declined with shorter scan durations without DL, whereas diagnostic confidence only negligibly changed with DL (intermediate-to-high confidence ratings: 0%-54% vs. 80%-84%; 83% for the 10-min scan). The diagnostic accuracy of PET reads dropped from 90% to 4% without and remained high with DL (90%-93%; 90% for the 10-min scan). <b>Conclusion:</b> Our study demonstrates the compelling performance of DL to restore cerebral [¹⁸F]FDG PET datasets with ultra-low-count statistics for quantitative regional, voxel-based group, and clinical visual analyses. Consequently, DL enables a dramatic reduction of scan durations or administered activities (e.g., 10-min scan with 3.5 MBq, equivalent to ∼60 µSv) for [¹⁸F]FDG PET in patients with dementia and possibly other indications.