Ultrafast Multi-tracer Total-body PET Imaging Using a Transformer-Based Deep Learning Model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Minimizing motion artifacts and patient discomfort in PET scans necessitates reducing acquisition time.
• Ultrafast PET imaging presents challenges in maintaining diagnostic image quality.

Purpose of the Study:

• To develop and evaluate a deep learning framework for synthesizing high-quality PET images from ultrafast scans.
• To assess the framework's performance in multi-tracer total-body PET imaging.

Main Methods:

• Retrospective analysis of clinical uEXPLORER PET/CT datasets ([18F]FDG, [18F]FAPI, [68Ga]FAPI).
• Generation of ultrafast PET images (3-40s) via list-mode data truncation.
• Development of two 3D SwinUNETR-V2 variants (PET-only and PET+CT fusion).
• 5-fold cross-validation for training and testing.

Main Results:

• Both deep learning models significantly improved subjective image quality and lesion detectability compared to original ultrafast scans.
• Objective image quality metrics, including peak signal-to-noise ratio (PSNR), were enhanced across ultra-short acquisitions.
• Significant improvements in PSNR were observed for [18F]FDG datasets at 3s, 6s, and 15s acquisition times (p < 0.001).

Conclusions:

• The proposed deep learning models effectively enhance image quality in multi-tracer total-body PET scans acquired with ultrafast protocols.
• Synthesized PET images demonstrate comparable performance to standard scans in terms of image quality and lesion detectability.
• This approach holds promise for faster and more comfortable PET imaging without compromising diagnostic accuracy.