Self-normalization for a 1 mm³resolution clinical PET system using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Nuclear Medicine

Background:

• Positron Emission Tomography (PET) imaging requires normalization for accurate quantitative analysis.
• Traditional normalization methods often necessitate separate, dedicated scans, adding complexity and time.
• Deep learning offers potential for image-based solutions to streamline PET data processing.

Purpose of the Study:

• To propose and evaluate an image-based, end-to-end self-normalization framework for PET.
• To leverage conditional generative adversarial networks (cGANs) for automated PET image normalization.
• To assess the impact of different input data types, network architectures, and input tensor shapes on normalization performance.

Main Methods:

• Development of a novel polarized self-attention (PSA) Pix2Pix deep learning network.
• Exploration of unnormalized vs. geometric-factors-corrected input images.
• Comparison of 2D vs. 2.5D input tensor shapes using Monte Carlo simulations (SimSET) with voxelized phantoms.

Main Results:

• The 2.5D PSA Pix2Pix model, using geometric-factors-corrected input, achieved superior performance.
• All tested methods improved image quality metrics (PSNR, SSIM) by 15-55%.
• The best approach yielded the highest PSNR (28.074) and SSIM (0.921) for overall images and (28.920, 0.973) for regions of interest, enhancing lesion detectability.

Conclusions:

• An image-based end-to-end self-normalization framework using cGANs is feasible for PET.
• The proposed 2.5D PSA Pix2Pix model significantly improves PET image quality and lesion detectability.
• This approach eliminates the need for separate normalization scans, offering a more efficient workflow.