Accelerating Clinical PET Image Denoising with Deep Learning Using NPS-Based Noise Simulation and Optimized Training Strategy

Purpose : To develop and evaluate a deep learning (DL) strategy for PET image denoising using scanner-specific noise power spectrum (NPS) simulations, enabling high-quality clinical reconstruction from low-dose or shortened acquisitions. Methods : A perceptually optimized 2.5D U-Net architecture was trained on paired clinical-clean and synthetically noised PET images, where the noise was generated by injecting NPS-derived correlated noise into clinical PET datasets from four scanners. Models were trained using datasets simulated with NPS (DL-NPS), list-mode reconstructions (DL-LM), or both (DL-all). The NPS dataset was augmented until denoising performance reached perceptually comparable PET images relative to LM-trained models, regarded as equivalent. Image quality was assessed using SSIM, PSNR, RMSE, liver coefficient of variation (CV), and lesion contrast-to-noise ratio (CNR). Clinical detectability was analyzed using the Rose Criterion and Cochran’s Q test. SUV agreement was assessed via Bland-Altman analysis. Results: The DL-NPS model performed comparably to DL-LM, leveraged by augmented data, while DL-all yielded further improved performance in specific scanner settings. At 25% and 50% counts, DL denoising improved CNR by 15–40% across systems, improving lesion detectability by up to 18%. Liver CV dropped below clinical thresholds post-denoising. Bland–Altman analysis showed minimal SUV bias (mean bias <0.1, 95% CI within ±1.2). Conclusion : NPS-based simulation enables realistic, scanner-specific noise generation without requiring gated or list-mode data. The proposed DL approach consistently restores image quality and lesion detectability across scanners and imaging settings. Its implementation supports scalable clinical adoption of low-dose PET protocols.