3D Wavelet-Based Structural Priors for Controlled Diffusion in Whole-Body Low-Dose PET Denoising

TL;DR

This work addresses the challenge of denoising whole-body ultra-low-dose PET while preserving anatomical fidelity. It introduces Wavelet-Conditioned ControlNet (WCC-Net), a fully 3D diffusion framework that injects a frequency-domain structural prior derived from a 3D Discrete Wavelet Transform into a frozen diffusion backbone via a ControlNet-style conditioning branch. The method achieves consistent improvements over CNN-, GAN-, and diffusion-based baselines, with strong gains in PSNR, SSIM, and structural fidelity on internal and unseen-dose evaluations, demonstrating robust anatomy-aware denoising in challenging noise regimes. The approach offers a practical path toward dose reduction in clinical PET by decoupling structural guidance from stochastic denoising, enabling more reliable quantitative and diagnostic outcomes.

Abstract

Low-dose Positron Emission Tomography (PET) imaging reduces patient radiation exposure but suffers from increased noise that degrades image quality and diagnostic reliability. Although diffusion models have demonstrated strong denoising capability, their stochastic nature makes it challenging to enforce anatomically consistent structures, particularly in low signal-to-noise regimes and volumetric whole-body imaging. We propose Wavelet-Conditioned ControlNet (WCC-Net), a fully 3D diffusion-based framework that introduces explicit frequency-domain structural priors via wavelet representations to guide volumetric PET denoising. By injecting wavelet-based structural guidance into a frozen pretrained diffusion backbone through a lightweight control branch, WCC-Net decouples anatomical structure from noise while preserving generative expressiveness and 3D structural continuity. Extensive experiments demonstrate that WCC-Net consistently outperforms CNN-, GAN-, and diffusion-based baselines. On the internal 1/20-dose test set, WCC-Net improves PSNR by +1.21 dB and SSIM by +0.008 over a strong diffusion baseline, while reducing structural distortion (GMSD) and intensity error (NMAE). Moreover, WCC-Net generalizes robustly to unseen dose levels (1/50 and 1/4), achieving superior quantitative performance and improved volumetric anatomical consistency.

Figures (4)

• Figure 1: Overview of the proposed Wavelet-Conditioned ControlNet (WCC-Net). (a) Diffusion backbone for low-dose to normal-dose PET patch denoising. (b) WCC-Net extracts multi-scale frequency structural priors via a 3D Discrete Wavelet Transform and injects them as conditional guidance into a frozen diffusion backbone, enabling fine-grained structure-aware and anatomically consistent PET denoising.
• Figure 2: Qualitative comparison of whole-body PET denoising results. Sagittal, coronal, and axial views are shown for a representative test subject, comparing the 1/20-dose input, competing denoising methods, and the normal-dose reference. Zoom-in regions (red boxes) highlight areas of high uptake and fine anatomical structures.
• Figure 3: Qualitative comparison of signed error maps for whole-body PET denoising at the 1/20 dose level. Sagittal, coronal, and axial views are shown for a representative test subject. Error maps are computed as the voxel-wise difference between each method and the normal-dose reference, and displayed using a diverging color scale centered at zero, where red and blue indicate over- and under-estimation, respectively.
• Figure 4: Intensity profile comparison along representative sagittal, coronal, and axial directions for the same test subject at the 1/20 dose level. Line profiles are extracted along the red dashed lines indicated in the corresponding PET slices. Zoom-in regions highlight local differences in peak preservation and structural transitions.