Deep supervised transformer-based noise-aware network for low-dose PET denoising across varying count levels

Area of Science:

• Medical Imaging
• Deep Learning
• Radiology

Background:

• Reducing radiation dose in Positron Emission Tomography (PET) imaging is crucial for minimizing cancer risks.
• Low-dose PET imaging often results in increased noise and degraded image quality, impacting diagnostic accuracy.
• Existing deep learning denoising methods struggle to generalize across diverse noise levels common in PET scans.

Purpose of the Study:

• To develop a unified deep learning network capable of handling diverse noise levels in low-dose PET imaging.
• To reconstruct high-quality PET images from low-dose acquisitions using a Swin Transformer-based approach.
• To overcome the limitations of existing models that require training for specific noise levels.

Main Methods:

• A Swin Transformer-based Noise-Aware Network (ST-UNN) was developed, integrating multiple sub-networks for noise levels from 1% to 10%.
• An adaptive weighting mechanism was employed to dynamically combine outputs from sub-networks for effective denoising.
• The model was trained and validated on a PET/CT dataset of the head and neck region, assessing performance with SSIM, PSNR, SUV bias, and RMSE.

Main Results:

• ST-UNN demonstrated superior performance compared to conventional networks, especially in ultra-low-dose scenarios (1% count level).
• At 1% count level, ST-UNN achieved a PSNR of 34.77, RMSE of 0.05, and SSIM of 0.97, with minimal SUV bias.
• The network maintained high performance and low error across all tested noise levels, indicating strong generalization and diagnostic integrity.

Conclusions:

• ST-UNN provides a scalable, transformer-based solution for denoising low-dose PET images.
• The dynamic integration of sub-networks effectively addresses noise variability, significantly improving image quality.
• This approach advances the capabilities of low-dose and dynamic PET imaging, enhancing diagnostic reliability.