Metabolically Faithful 3D PET Restoration via Volumetric Swin Transformers

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Positron Emission Tomography (PET) diagnostic precision is limited by low spatial resolution.
• Current deep learning models often trade quantitative accuracy for visual sharpness and lack generalization due to fixed training degradations.

Purpose of the Study:

• To develop a metabolically faithful 3D restoration framework for PET imaging.
• To improve spatial resolution and quantitative accuracy in PET scans.

Main Methods:

• A volumetric extension of SwinFIR was paired with a composite metabolic-aware loss function.
• Stochastic degradation augmentation was employed to simulate diverse scanner-like degradations during training.

Main Results:

• The proposed method achieved superior performance with SSIM of 0.843, PSNR of 27.08 dB, and NRMSE of 0.117.
• Metabolic fidelity was maintained, indicated by a CCC of 0.948 and Wasserstein distance of 0.018.
• Stochastic augmentation enhanced robustness compared to fixed-profile training.

Conclusions:

• The framework successfully recovers anatomical detail in PET images.
• Small, symmetric regional SUVR biases (≤3.3%) were observed in cortical regions across diagnostic groups.
• The developed method offers improved PET image restoration with preserved metabolic accuracy.