Clinical utility of deep learning for the recovery of standard-dose imaging quality from low-dose PET

Ziel/Aim Even with radiation dose of modern PET imaging becoming lower over time, it remains a bottleneck for this modality’s extensive application. Artificial intelligence has recently been developed to recover high-quality imaging from low-dose scans but generalizing and applying this clinically is challenging.

Methodik/Methods Brain FDG-PET of 237 patients from scanner 1 (digital, vendor 1) was used to develop AI technology. The AI algorithm was used on FDG-PET of patients with suspected neurodegenerative disease from 2 scanners (analogue, vendor 2, n = 20 and digital, vendor 2, n = 7). Original dose images (OD) and images mimicking a dose reduction fraction (DRF) of up to 50 (AI-enhanced/non-enhanced) were reconstructed. Two nuclear medicine physicians assessed the comparability of the data in a clinical setting e.g. by rating the hypometabolism (4-point scale, 0=no hypometabolism to 3=strong hypometabolism) on 3D-SSP analysis of OD and DRF 50 data (AI-enhanced/non-enhanced). Results were compared between groups by Friedman test (p < 0.05) with post-hoc tests (Wilcoxon signed-rank test + Bonferroni adjustment).

Ergebnisse/Results The quantitative assessment of the data showed an advantage of AI-enhancement for NRMSE, SSIM and PSNR. For rater 1 post-hoc tests showed significant differences between OD and DRF 50 non-enhanced (p 0.000) but not OD and DRF 50 AI-enhanced data. For rater 2 the Friedman test showed no significant difference (p 0.202) between groups. AI-enhanced data had a tendency to be better and closer to the OD on visual assessment of axial images.

Schlussfolgerungen/Conclusions The developed AI method for low-dose PET image enhancement can be generalized for different scanners. Clinical evaluation showed a tendency for the AI to be advantageous. Results for rater 2 showing no significant difference between the different groups indicates that even high DRF non-AI data has good quality. Therefore, using even higher DRFs or injecting lower doses regardless of the use of AI might be possible in new PET/CT scanners.

Publication History

Article published online:
08 April 2021