PET Image Reconstruction Using Deep Diffusion Image Prior

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.

💡 Research Summary

This paper introduces a novel PET image reconstruction framework that leverages diffusion models as powerful priors while addressing two major challenges: (1) the variability of tracer‑specific contrast and scanner differences that cause out‑of‑distribution (OOD) performance degradation, and (2) the high computational burden of fully‑3D diffusion‑based reconstruction. Inspired by the Deep Diffusion Image Prior (DDIP) concept, the authors propose an alternating scheme that interleaves diffusion sampling with sinogram‑driven fine‑tuning of a pretrained score network. The method incorporates an anatomical prior (e.g., patient‑specific MR) to condition the reverse diffusion process, thereby guiding reconstruction toward anatomically plausible solutions.

Key technical components are:

1. Conditional Diffusion Model – The reverse process pθ(x_{t‑1}|x_t,g) is modeled with a U‑Net‑based score function ϵθ(x_t,t,g). The anatomical prior g is supplied at each timestep, enabling patient‑specific conditioning.

2. Half‑Quadratic Splitting (HQS) – To decouple the large‑scale linear PET forward model (system matrix A) from the deep network, the authors formulate a joint optimization over the image x and network parameters θ. HQS splits this into (a) an image update solved via a surrogate Poisson log‑likelihood derived from MLEM, and (b) a network update solved by standard gradient descent. This alternating minimization yields closed‑form voxel‑wise updates (Eq. 19) and dramatically reduces memory consumption.

3. Low‑Rank Adaptation (LoRA) – Instead of fine‑tuning all network weights, only low‑rank factors ΔW = UV are learned, where U∈ℝ^{d×r} and V∈ℝ^{r×k} with r≪min(d,k). In experiments r=4, representing only ~1 % of total parameters, preserving the pretrained knowledge while allowing rapid adaptation to new tracer distributions.

4. Deterministic Sampling (DDIM) with Stochasticity Control – The reverse diffusion step uses the DDIM update (Eq. 10) with a controllable stochasticity parameter η∈