Fed-NDIF: A Noise-Embedded Federated Diffusion Model For Low-Count Whole-Body PET Denoising

TL;DR

This work tackles denoising of low-count PET while preserving patient privacy by introducing Fed-NDIF, a noise-embedded federated diffusion framework. It encodes count-level noise through liver $NSTD$ and trains a 2.5D conditional diffusion model using FedAvg across three multicenter datasets, followed by local fine-tuning. Fed-NDIF achieves superior image quality and lesion quantification (PSNR, SSIM, NMSE, SUV metrics) compared with federated UNet baselines and local diffusion models across multiple count levels. The study demonstrates that combining diffusion models with privacy-preserving federated learning and noise-aware conditioning effectively handles cross-site noise variability, offering a practical path toward safer low-dose PET imaging.

Abstract

Low-count positron emission tomography (LCPET) imaging can reduce patients' exposure to radiation but often suffers from increased image noise and reduced lesion detectability, necessitating effective denoising techniques. Diffusion models have shown promise in LCPET denoising for recovering degraded image quality. However, training such models requires large and diverse datasets, which are challenging to obtain in the medical domain. To address data scarcity and privacy concerns, we combine diffusion models with federated learning -- a decentralized training approach where models are trained individually at different sites, and their parameters are aggregated on a central server over multiple iterations. The variation in scanner types and image noise levels within and across institutions poses additional challenges for federated learning in LCPET denoising. In this study, we propose a novel noise-embedded federated learning diffusion model (Fed-NDIF) to address these challenges, leveraging a multicenter dataset and varying count levels. Our approach incorporates liver normalized standard deviation (NSTD) noise embedding into a 2.5D diffusion model and utilizes the Federated Averaging (FedAvg) algorithm to aggregate locally trained models into a global model, which is subsequently fine-tuned on local datasets to optimize performance and obtain personalized models. Extensive validation on datasets from the University of Bern, Ruijin Hospital in Shanghai, and Yale-New Haven Hospital demonstrates the superior performance of our method in enhancing image quality and improving lesion quantification. The Fed-NDIF model shows significant improvements in PSNR, SSIM, and NMSE of the entire 3D volume, as well as enhanced lesion detectability and quantification, compared to local diffusion models and federated UNet-based models.

Figures (8)

• Figure 1: The image noise level is determined by NSTD of a 2 cm x 2 cm x 2 cm cube within the liver region. The liver NSTD distribution for each count level across different datasets is shown on the left. On the right, three example images from each dataset are displayed, illustrating the variation in image noise across different count levels and datasets.
• Figure 2: Overview of the proposed Fed-NDIF. In Stage A, the liver NSTD on LCPET is calculated and later embedded into the diffusion model in Stage B. During Stage B, local diffusion models are trained independently at each site. Aggregation occurs every few epochs, wherein model parameters are averaged to update the centralized model. This centralized model is then sent back to the local sites to initiate a new round of local training. In Stage C, the global model is further fine-tuned using local datasets.
• Figure 3: Sample images from three institutions with the lowest count levels are shown. The yellow boxes are cropped regions that are magnified and displayed at the bottom-right corner of the image. The blue arrows indicate the lesions on FCPET. Federated diffusion models produce images with higher resolution and fewer false-positive and false-negative lesions, as indicated by blue arrows and yellow boxes.
• Figure 4: Comparison of the local diffusion model and the federated diffusion model on 1%, 5%, 10% count levels for Bern data. Images are normalized to standard count. The left shows the denoised images, and the right shows the absolute error map of the two methods. Fed-NDIF has reduced the overestimation in liver and cardiac region and underestimation of the body.
• Figure 5: Comparison of the local diffusion model and the federated diffusion model on 1%, 5%, 10% count levels for Ruijin data. Images are normalized to standard count. Fed-NDIF has reduced the overestimation and underestimation of the denoised outputs.