SMART-PET: a Self-SiMilARiTy-aware generative adversarial framework for reconstructing low-count [18F]-FDG-PET brain...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Positron Emission Tomography (PET) imaging involves radioactive tracers, increasing radiation exposure, especially in pediatric and longitudinal studies.
• Reducing PET tracer activity compromises image quality (lower signal-to-noise ratio) and diagnostic accuracy.
• Current deep learning denoising methods often require anatomical guidance (e.g., MRI) and struggle to preserve global spatial features in PET images.

Purpose of the Study:

• To develop a novel PET-only deep learning framework for denoising low-count PET images.
• To reduce radiation exposure in PET imaging without sacrificing image quality.
• To create a framework applicable to radiosensitive populations and longitudinal studies.

Main Methods:

• Developed the Self-SiMilARiTy-Aware Generative Adversarial Framework (SMART), a PET-only deep learning model using Generative Adversarial Networks (GANs).
• Incorporated a self-similarity attention mechanism (SSAB) to learn distinctive features for denoising without MRI guidance.
• Trained the SMART GAN on a dataset of 114 subjects (epilepsy, dementia, healthy volunteers) using standard-dose PET images as reference.

Main Results:

• SMART-PET achieved high image quality metrics compared to standard-dose PET, including SSIM (0.984), PSNR (38.126 dB), and low NRMSE (0.091).
• The model demonstrated excellent performance with FID (0.455), SNR (0.002), and CNR (0.011).
• Region of interest measurements from 10% count datasets showed less than 1.4% deviation from ground-truth values.

Conclusions:

• SMART-PET effectively denoises PET images, synthesizing diagnostic quality scans with a 90% reduction in injected activity.
• The framework shows significant promise for clinical applications, particularly for radiosensitive patient groups and long-term neurological monitoring.
• This PET-only approach eliminates the need for co-registered MRI, simplifying the imaging protocol.