Dual-domain transformer-based learned primal-dual reconstruction for PET imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Positron Emission Tomography (PET) is crucial for metabolic assessment and cancer diagnosis.
• High noise in PET data necessitates advanced reconstruction algorithms for accuracy.
• Current Convolutional Neural Network (CNN)-based Learned Primal-Dual (LPD) methods struggle with long-range dependencies in sinogram data.

Purpose of the Study:

• To investigate the integration of attention mechanisms (self-attention and cross-attention) into the LPD framework for PET reconstruction.
• To develop and evaluate novel transformer-based LPD architectures for improved PET imaging.
• To create a robust synthetic data generation process for training and transfer learning.

Main Methods:

• Proposed three transformer-based LPD architectures: Dual-Domain Stacked Transformer-LPD, Dual-Domain Restormer-LPD, and Dual-Domain UNet-LPD with Cross-Attention.
• Developed a system-aligned synthetic data generation process mimicking a preclinical PET scanner (MiniPET-3).
• Incorporated realistic factors like geometry, blurring, positron range, and Poisson noise into synthetic data.

Main Results:

• Transformer-based LPD models (Stacked Transformer and Restormer) outperformed the UNet-LPD baseline in Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR).
• The Restormer model achieved higher Structural Similarity Index Measure (SSIM) at low noise levels.
• The UNet-LPD with cross-attention improved MSE and PSNR but showed a slight SSIM reduction, indicating a trade-off.
• Models trained on synthetic data generalized effectively to experimental measurements, yielding high-quality reconstructions.

Conclusions:

• Transformer-integrated LPD frameworks show significant promise for enhancing PET reconstruction.
• These advanced models improve image quality and robustness, particularly in low-count or noisy PET imaging scenarios.
• The developed synthetic data generation enables effective transfer learning for preclinical PET systems.