Whole-body PET image denoising for reduced acquisition time.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Reduced Positron Emission Tomography (PET) acquisition times enhance patient comfort and scanner efficiency.
• Artificial intelligence (AI) offers potential for accelerating PET scans while maintaining diagnostic image quality.

Purpose of the Study:

• To compare the efficacy of various neural network architectures for PET image denoising.
• To identify optimal AI models for reducing PET acquisition duration without compromising image fidelity.

Main Methods:

• Evaluated 2D/2.5D ResNet, Unet, SwinIR, and 3D MedNeXt, 3D UX-Net models on 212 studies (56,908 images) with 7MBq/kg injected activity.
• Utilized 2D metrics (RMSE, SSIM) and 3D metrics (SUVpeak, SUVmax error in ROIs) for model assessment.
• Compared supervised learning approaches against unsupervised CycleGAN for PET denoising.

Main Results:

• 3D MedNeXt demonstrated superior performance in PET denoising tasks.
• 3D MedNeXt achieved significant improvements in SSIM (38.2%) and RMSE (28.1%) for 30-s denoised PET scans compared to original 90-s scans.
• Further improvements were observed for 60-s denoised scans (SSIM 16.9%, RMSE 11.4%), with reduced SUVmax discrepancy dispersion.

Conclusions:

• 3D MedNeXt is the most effective AI model identified for PET denoising.
• AI-driven PET denoising, particularly with 3D MedNeXt, enables substantial reductions in scan time.
• This advancement promises enhanced patient experience and increased throughput in PET imaging centers.