Evaluation of PET/CT Artificial Intelligence Image Reconstructions VS Harmonized Clinical Reconstruction

Background This study focuses on evaluating how SubtlePET™, an artificial intelligence (AI)-based image enhancement algorithm, produced PET/CT images compare against EARL Standard-2 harmonized images, and to assess its impact on quantitative accuracy and workflow in a clinical setting. Methods A NEMA IEC PET Body Phantom with 10:1 (EARL protocol) and 4:1 (NEMA NU2-2018 protocol) sphere-to-background ratios were scanned using a General Electric Discovery™ IQ Gen-2 PET/CT system. Data were acquired in list mode for 10 minutes per bed position and retrospectively reconstructed to simulate reduced scan duration. Images were reconstructed with and without SubtlePET TM (phantom image trained) using EARL Standard-2 accredited reconstruction. Recovery coefficients (SUV mean , max , peak ) were calculated and compared to EARL standard-2 along with their repeatability. Additionally, 23 patient datasets acquired for 3-minutes per bed-position were reconstructed to simulate reduced acquisition time both by SubtlePET TM (human image trained) and EARL Standard-2 reconstruction parameters. Lesions were segmented semi-automatically to assess SUV and volume metrics. Statistical comparisons were performed using a two-sample t-test assuming equal variances. Results SubtlePET TM reconstructions achieved 91% compliance with EARL standard-2. SUV peak remained the most robust metric across scan duration. Patient data showed no significant differences in SUV peak or volume for SubtlePET TM at half the acquisition time. Minor deviations were noted in SUV max and SUV mean for smaller lesions at 1-min/bed. Conclusion SubtlePET™ preserves quantitative performance comparable to EARL Standard 2–compliant reconstructions, particularly when SUVpeak is used, although deviations associated with reduced acquisition time persist. These findings support the potential for reduced scan duration, which may facilitate lower radiation exposure or increased patient throughput without compromising diagnostic quality. Further validation across different scanners and clinical settings is warranted.