Ablation Study of Diffusion Model with Transformer Backbone for Low-count PET Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Restoration

Background:

• Diffusion models (DM) are advanced generative models increasingly applied to image restoration (IR).
• Vision transformers, like Restormer, have emerged as powerful backbones for IR tasks, evolving from UNet architectures.
• The efficacy of DMs, particularly transformer-based ones, for Positron Emission Tomography (PET) denoising remains under-explored.

Purpose of the Study:

• To investigate if diffusion models act as a general add-on generative learning scheme to boost PET denoising when using a powerful backbone.
• To disentangle the contributions of backbone networks versus generative learning schemes in PET denoising.
• To identify best practices for PET denoising by comparing different model architectures and diffusion strategies.

Main Methods:

• A latent diffusion model (DiffIR) based on the Restormer backbone was evaluated for ¹⁸F-FDG whole-body PET denoising.
• Comparisons were made against UNet, SR3 (UNet + pixel-space DM), and Restormer on low-dose (25%) PET data.
• The study involved training on 93 subjects and testing on 12 subjects, with 644 slices per subject.

Main Results:

• Restormer significantly outperformed UNet in denoising performance based on Peak Signal-to-Noise Ratio (PSNR) and Mean Squared Error (MSE).
• Adding latent diffusion to Restormer did not improve MSE, Structural Similarity Index Measure (SSIM), or PSNR, and was inferior to UNet.
• SR3 with pixel-space diffusion produced unstable and unsatisfactory results, similar to findings in natural image super-resolution tasks.

Conclusions:

• A powerful backbone like Restormer is crucial for PET denoising, outperforming simpler UNet architectures.
• Latent diffusion models may not be a universally beneficial add-on for PET denoising, potentially struggling with detailed structure and texture restoration.
• The limited spatial information in low-dose PET and super-resolution tasks poses challenges for diffusion model performance in restoring fine details.