Perceptually optimized NPS-based deep learning framework for PET image denoising.
Authors
Affiliations (6)
Affiliations (6)
- Klar Inteligência Artificial, Recife, PE, Brazil.
- Klar Inteligência Artificial, Recife, PE, Brazil; Hospital das Clínicas da Universidade Federal de Pernambuco (HC/UFPE-HU Brasil), Recife, PE, Brazil.
- Klar Inteligência Artificial, Recife, PE, Brazil; Hospital Universitário Professor Alberto Antunes (HUPAA-UFAL- HU Brasil), Maceió, AL, Brazil.
- Hospital das Clínicas da Universidade Federal de Pernambuco (HC/UFPE-HU Brasil), Recife, PE, Brazil.
- Centro de Informática, Universidade Federal de Pernambuco (CIn-UFPE), Recife, PE, Brazil.
- Klar Inteligência Artificial, Recife, PE, Brazil; Hospital Universitário Professor Edgard Santos (HUPES-UFBA- HU Brasil), Salvador, BA, Brazil. Electronic address: [email protected].
Abstract
To develop and evaluate a scanner-specific noise power spectrum (NPS)-based framework for deep learning PET image denoising under reduced-count conditions. Retrospective whole-body <sup>18</sup>F-FDG PET/CT datasets from 250 subjects and three scanners were used to train a perceptually optimized 2.5D U-Net. Data were partitioned at the patient level into scanner-specific training, validation, and independent testing subsets. Training strategies included NPS-simulated data (DL-NPS), list-mode reconstructions (DL-LM), and their combination (DL-all). Image quality was evaluated using SSIM, PSNR, RMSE, liver coefficient of variation (CV), lesion contrast-to-noise ratio (CNR), and SUV agreement. CNR-based lesion visibility was assessed using the Rose Criterion, while three nuclear medicine physicians independently performed blinded image-quality assessments. DL-NPS achieved competitive performance relative to LM-trained models across scanner configurations. Denoising improved lesion CNR by 15-38% at 25% and 50% counts, increasing the proportion of lesions meeting the Rose visibility threshold by up to 13.5% after denoising. Liver CV decreased toward standard-count reference levels. Bland-Altman analysis showed limited SUV bias (absolute mean bias ≤0.17, 95% LoA within ±1.1). Reader assessments generally favored DL-denoised images, with significant paired improvements for Biograph Horizon but not Discovery 710 after multiplicity correction. Scanner-specific NPS simulation provides a feasible strategy for generating PET denoising training data when LM acquisitions are limited. Larger independent and external studies are required to establish generalizability and clinical applicability.