PET Image Reconstruction Using Deep Diffusion Image Prior

TL;DR

This work addresses the challenge of reconstructing high-quality PET images under low-dose conditions and tracer/ scanner variability. It introduces a diffusion model–based framework (DDIP) guided by anatomical priors, where diffusion sampling is alternated with score-function fine-tuning and decoupled from the PET forward model via half-quadratic splitting to improve efficiency. The approach demonstrates robust out-of-distribution generalization across tracers and scanners, enabling reliable reconstruction with reduced data and improved structural fidelity, as shown in both simulations and clinical datasets. Overall, the method offers a versatile, computationally efficient path toward cross-tracer, low-dose PET reconstruction with strong anatomical consistency.

Abstract

Diffusion models have shown great promise in medical image denoising and reconstruction, but their application to Positron Emission Tomography (PET) imaging remains limited by tracer-specific contrast variability and high computational demands. In this work, we proposed an anatomical prior-guided PET image reconstruction method based on diffusion models, inspired by the deep diffusion image prior (DDIP) framework. The proposed method alternated between diffusion sampling and model fine-tuning guided by the PET sinogram, enabling the reconstruction of high-quality images from various PET tracers using a score function pretrained on a dataset of another tracer. To improve computational efficiency, the half-quadratic splitting (HQS) algorithm was adopted to decouple network optimization from iterative PET reconstruction. The proposed method was evaluated using one simulation and two clinical datasets. For the simulation study, a model pretrained on [$^{18}$F]FDG data was tested on [$^{18}$F]FDG data and amyloid-negative PET data to assess out-of-distribution (OOD) performance. For the clinical-data validation, ten low-dose [$^{18}$F]FDG datasets and one [$^{18}$F]Florbetapir dataset were tested on a model pretrained on data from another tracer. Experiment results show that the proposed PET reconstruction method can generalize robustly across tracer distributions and scanner types, providing an efficient and versatile reconstruction framework for low-dose PET imaging.

Figures (12)

• Figure 1: Overview of the proposed DDIP-based and anatomical prior-guided PET image reconstruction framework. At each time step $t$, the proposed method comprises two sub-steps: (a) a fine-tuning sub-step, solved using the HQS; and (b) a DDIM sampling sub-step.
• Figure 2: Ablation study on the combinations of components for DDIP.
• Figure 3: Simulation results of amyloid-negative data under different LoRA configurations, including varying rank values $r$.
• Figure 4: Simulation results of the amyloid-negative data for different starting times $T'$. Rows correspond to the reconstructed images (top) and the error maps (bottom), where each error map is defined as target image - ground truth. The grayscale bar at the top indicates activity (a.u.); and the bottom color bar indicates signed differences (a.u.).
• Figure 5: Effect of the hyperparameter $\beta$ on PSNR for the simulation data.