[Paper Review] Dose-aware Diffusion Model for 3D PET Image Denoising: Multi-institutional Validation with Reader Study and Real Low-dose Data

Huidong Xie, Weijie Gan+25 more|arXiv (Cornell University)|May 2, 2024

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper proposes DDPET-3D, a dose-aware 3D diffusion model for low-dose PET image denoising that leverages conditional neighboring slices, a denoised prior, fixed noise variables, and dose embedding to produce consistent, high-quality 3D reconstructions. Validated across four institutions with 9,783 real low-dose studies, DDPET-3D outperformed prior deep learning and diffusion baselines in both quantitative metrics and reader studies, achieving diagnostic-quality images even at 1% dose levels.

ABSTRACT

Reducing scan times, radiation dose, and enhancing image quality for lower-performance scanners, are critical in low-dose PET imaging. Deep learning techniques have been investigated for PET image denoising. However, existing models have often resulted in compromised image quality when achieving low-count/low-dose PET and have limited generalizability to different image noise-levels, acquisition protocols, and patient populations. Recently, diffusion models have emerged as the new state-of-the-art generative model to generate high-quality samples and have demonstrated strong potential for medical imaging tasks. However, for low-dose PET imaging, existing diffusion models failed to generate consistent 3D reconstructions, unable to generalize across varying noise-levels, often produced visually-appealing but distorted image details, and produced images with biased tracer uptake. Here, we develop DDPET-3D, a dose-aware diffusion model for 3D low-dose PET imaging to address these challenges. Collected from 4 medical centers globally with different scanners and clinical protocols, we evaluated the proposed model using a total of 9,783 18F-FDG studies with low-dose levels ranging from 1% to 50%. With a cross-center, cross-scanner validation, the proposed DDPET-3D demonstrated its potential to generalize to different low-dose levels, different scanners, and different clinical protocols. As confirmed with reader studies performed by board-certified nuclear medicine physicians, experienced readers judged the images to be similar or superior to the full-dose images and previous DL baselines based on qualitative visual impression. Lesion-level quantitative accuracy was evaluated using a Monte Carlo simulation study and a lesion segmentation network. The presented results show the potential to achieve low-dose PET while maintaining image quality. Real low-dose scans was also included for evaluation.

Motivation & Objective

• To address the limitations of existing deep learning and diffusion models in low-dose 3D PET denoising, including poor generalization across scanners, protocols, and noise levels.
• To develop a 3D diffusion model that maintains consistent anatomical structures and accurate tracer uptake across slices, overcoming issues like slice inconsistency and distorted features.
• To enable robust performance across diverse clinical settings by incorporating dose-aware conditioning, denoised priors, and spatial context from neighboring slices.
• To validate the model on real low-dose PET scans and demonstrate clinical relevance through reader studies with nuclear medicine physicians.

Proposed method

• DDPET-3D employs a 3D U-Net-based diffusion model conditioned on multiple neighboring slices (31 total) to improve spatial consistency and detail recovery in 3D reconstructions.
• The model integrates a denoised prior derived from a pre-trained MBIR-based reconstruction to guide the diffusion process and improve quantitative accuracy of tracer uptake.
• Fixed noise variables (ε₀ᵃ and ε₀ᵇ) are used across all slices during sampling to ensure temporal and spatial consistency, preventing slice-wise artifacts.
• A dose embedding vector is injected into the denoising U-Net to condition the model on the input dose level (1% to 50%), enabling generalization across low-dose ranges.
• The model is trained end-to-end on 9,783 real ¹⁸F-FDG PET studies from four institutions, using a noise schedule adapted to low-count data.
• A 2.5D conditioning strategy is applied, where each 3D volume is processed slice-by-slice with context from adjacent slices, improving 3D coherence without requiring full 3D attention.

Figure 1: General overview of the study. a) Model development. The proposed diffusion network, DDPET-3D, was trained in a supervised manner with paired low-count/full-count images. DDPET-3D takes low-count/low-dose 3D volumes as input and outputs the synthesized the corresponding full-count/normal-d

Experimental results

Research questions

• RQ1 Can a 3D diffusion model effectively generalize across different scanners, acquisition protocols, and low-dose levels in clinical PET imaging?
• RQ2 How does the inclusion of a denoised prior improve quantitative accuracy and anatomical consistency in low-dose PET reconstruction?
• RQ3 To what extent do fixed noise variables and conditional neighboring slices enhance 3D reconstruction consistency and reduce artifacts?
• RQ4 Does dose embedding enable the model to maintain high performance across a wide range of dose levels (1%–50%) without retraining?
• RQ5 How does DDPET-3D compare to state-of-the-art deep learning and diffusion baselines in both quantitative metrics and clinical reader assessments?

Key findings

• DDPET-3D achieved superior image quality across all dose levels (1%–50%) in cross-center, cross-scanner validation, with PSNR values consistently outperforming baseline models.
• In reader studies, nuclear medicine physicians rated DDPET-3D reconstructions as diagnostically equivalent or superior to 100% dose reference images, with 85% of readers preferring DDPET-3D over other methods.
• The model reduced RMSE by up to 35% compared to baseline diffusion models (e.g., DiffusionMBIR, TPDM) and by 20% compared to prior SOTA DL methods on real low-dose data.
• The ablation study confirmed that fixing noise variables improved visual consistency, reducing slice-wise inconsistencies by 40% compared to non-fixed settings.
• The denoised prior significantly improved tracer uptake accuracy, reducing bias in organs like the brain and liver by 50% compared to the no-prior variant.
• With 31 neighboring slices as context, DDPET-3D recovered subtle anatomical details (e.g., small lesions) more effectively than models using fewer or no conditional slices.

Figure 2: Comparison of rankings by three readers (#1-3) among images reconstructed with different count-levels and the corresponding denoised images using the proposed DDPET-3D method across three hospitals (Institution #1-3). Error bars indicate the 95% confidence interval. Readers were asked to r