RT Journal Article SR Electronic T1 Low-count Rb-82 cardiac PET/CT image denoising via deep learning for clinical dose reduction with preserved quantification accuracy. JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 2263 OP 2263 VO 63 IS supplement 2 A1 Fridlund, Michelle A1 Lassen, Martin A1 Hasbak, Philip A1 Højgaard, Liselotte A1 Munk, Ole A1 Andersen, Flemming A1 Ladefoged, Claes YR 2022 UL http://jnm.snmjournals.org/content/63/supplement_2/2263.abstract AB 2263 Introduction: Myocardial perfusion imaging (MPI) with PET allows for evaluation of perfusion defect sizes. Current guidelines recommend injections of 1100-1500 MBq (30-40 mCi) Rubidium-82 (82Rb) for MPI. Such thresholds might saturate the PET/CT systems; hence reductions in the administered doses are desired. However, directly limiting the amount of tracer leads to impaired image quality in the form of increased noise. The purpose of this study was to develop a robust AI deep learning (DL) approach to obtain standard-dose image quality from quarter-dose reconstructions without compromising the clinically relevant quantification accuracy and allowing for dose reduction in future studies. Methods: This retrospective study comprised 82Rb MPI from 111 patients with known or suspected cardiac disease, imaged on a Siemens Biograph 128 mCT/PET scanner at rest and adenosine-induced stress states from June 2016 to July 2020. Only patients who received between 1000 MBq and 1200 MBq of 82Rb were included in this study. The resulting clinical listmode data were decimated with the Siemens e7 tools to simulate their respective quarter-dose counterparts. Images at both dose levels were reconstructed with the PSF-TOF 3D OSEM algorithm, using 4 iterations, 21 subsets, matrix size of 128×128×111 and a 2 mm Gaussian postfiltering. The high-/low-dose image pairs were then used to train a DL algorithm. 84 patients were reserved for training, 17 for validation and 10 for clinical metric testing. The model was validated by comparison of quantitative image metrics peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and normalized root-mean-squared-error (NRMSE). A paired t-test was used to verify the noise reduction outcomes. Total perfusion deficit (defect extent and severity) values at rest and stress (rTPD/sTPD) in the left ventricle (LV) were extracted with the Cedars QPS software. Correlation between TPD values before and after denoising was compared with Pearson’s R coefficient.Results: The AI-denoising yielded quantitative improvement across all images (PSNR: 37.53 ± 0.40/41.52 ± 0.30, SSIM: 0.89 ± 0.01/0.95 ± 0.00, NRMSE: 1.12 ± 0.04/0.70 ± 0.02 in input and output respectively; expressed as mean ± SD). All relative differences between the means were statistically significant (p < 0.0001). Substantial noise reduction was observed in DL-enhanced quarter-dose images (Figure 1 shows a single slice example from one of the test patients). Out of the 10 patients selected for testing 5 exhibited normal TPD (< 5% of LV), 2 showed moderate (10-15%) and 3 severe (> 20%) abnormalities during the stress acquisition. Only the 3 patients with severe abnormalities had fixed perfusion defects, present during rest test (between 5 and 15% of LV), while the remainder had little to none. The denoising method did not introduce any incongruencies to neither rest nor stress TPDs. Their values pre- and post-denoising resembled those of the reference dose, with strong correlation in both (RsTPD: 0.99/0.99; p < 0.0001 and RrTPD: 0.98/0.98; p < 0.0001). The regression slope was 1.01 for sTPD before and after denoising, while a slight underestimation was observed in denoised rTPD values (0.99 and 0.85 respectively).Conclusions: The proposed AI method demonstrated excellent denoising ability without significant loss of diagnostic accuracy for low-dose 82Rb cardiac PET/CT images. Dose reduction to 25% with quantitative clinical metrics comparable to reference dose was possible.