Clinical value of artificial intelligence in reducing PET image acquisition time: routine clinical validation using qualitative, quantitative, and radiomic analysis on a cohort of 282 patients undergoing [<sup>18</sup>F]FDG and [<sup>68</sup>Ga]Ga-PSMA-11 PET/CT.
Authors
Affiliations (5)
Affiliations (5)
- CRCI2NA- Inserm UMR1307/CNRS UMR 6075, University of Angers, University Hospital of Angers, Angers, France. [email protected].
- Department of nuclear medicine, Paul Papin Center, Institut de Cancérologie de l'Ouest (ICO), Angers, France.
- CRCI2NA- Inserm UMR1307/CNRS UMR 6075, University of Angers, University Hospital of Angers, Angers, France.
- Department of nuclear medicine, University Hospital of Angers, Angers, France.
- Department of Medical Physics, Paul Papin Center, Institut de Cancérologie de l'Ouest (ICO), Angers, France.
Abstract
Reducing acquisition time in PET/CT imaging can degrade image quality and may compromise both diagnostic reliability and the robustness of radiomic features. This study investigates, in a large clinical cohort, whether AI-based denoising can preserve image quality and maintain the accuracy of quantitative and radiomic parameters in [<sup>18</sup>F]FDG and [<sup>68</sup>Ga]Ga-PSMA-11 PET/CT scans. We reconstructed three sets of images: 100% acquisition time (R100), 75% (R75), and 50% (R50), with their respective denoised versions using SubtlePET® (S75 and S50). On a NEMA phantom, we analysed six contrasts (12:1-2:1) using [<sup>18</sup>F]FDG, assessing contrast, background noise, and radiomic features. In a cohort of 282 patients injected with [<sup>18</sup>F]FDG and [<sup>68</sup>Ga]Ga-PSMA-11, five nuclear medicine physicians performed a qualitative evaluation of image quality and confidence in the presence of hypermetabolism in 634 lesions. 109 radiomic features from 105 lesions were compared between the original and denoised reconstructions. The phantom study showed no difference in sphere contrast, a reduction in background noise variability, and excellent preservation of radiomic features. In the clinical population, S75 images showed improvements across all criteria evaluated, except for diagnostic confidence, which remained higher with R75 (p = 0.555 when compared to R100) for [<sup>18</sup>F]FDG. For [<sup>68</sup>Ga]Ga-PSMA-11, only S50 images showed a significant degradation in liver image quality. A decrease in SUVmax was observed in denoised images (- 7.73% for [<sup>18</sup>F]FDG; - 11.46% for [<sup>68</sup>Ga]Ga-PSMA-11, p < 0.0001). The radiomic analysis demonstrated excellent correlation, with a concordance correlation coefficient (CCC) > 0.8 for 90% of radiomic features. SubtlePET® improves the image quality of PET acquisitions performed with reduced acquisition times using [<sup>18</sup>F]FDG and [<sup>68</sup>Ga]Ga-PSMA. However, clinician confidence remains limited and, while denoised acquisitions preserve most radiomic features, others are altered, potentially limiting model transposability.