Imaging quality of an AI denoising algorithm: Validation in 68Ga PSMA-11 PET of patients with biochemical recurrence of prostate cancer

68Ga-PSMA PET imaging is the leading prostate cancer tracer but image quality remains noisy and could be improved by an AI-based denoising algorithm. The objective was to analyze the overall quality of reprocessed images in comparison with standard reconstructions. We also analysed the diagnostic performances of the different sequences and the impatch of the algorithm on lesion intensity and background measures. We retrospectively included 30 patients with biochemical recurrence of prostate cancer who performed a 68Ga-PSMA-11 PET-CT. We simulated images using only a quarter, half, three quarters or all the acquired data materials reprocessed by the SubtlePET® denoising algorithm. Three physicians with different levels of experience blindly analysed every sequence and a 5-level Likert scale was used to assess the series. The binary criterion of lesion detectability was compared between series. We also compared the SUV of lesions, background noise, and the diagnostic performances of the series (sensitivity, specificity, accuracy). Regarding image quality, the VPFX-derived series were classified differently but better than the standard (P < 0.001) from half the data. For QClear series, there was no difference for series from half of the signal. Although some series were noisy, lesion detectability did not differ (P > 0.05). Regarding the influence of SubtlePET® on SUV measurements, the algorithm significantly decreased lesion values (P < 0.005) and increased liver background (P < 0.005). The diagnostic performance of each reader was little changed by the algorithm. The AI algorithm could therefore be used for 68Ga-PSMA examinations using half the signal with similar image quality for QClear series and superior quality for VPFX series. However, this algorithm significantly modifies the quantitative measurements and should not be used for comparative examinations in case of reconstructed anteriorities with standard algorithm.