Multi-Subject Image Synthesis as a Generative Prior for Single-Subject PET Image Reconstruction

TL;DR

PET reconstruction quality is limited by Poisson noise and finite counts. The authors generate a large, diverse set of pseudo-PET images by MR-guided deformable registration of multi-subject PET-MR pairs, warping each PET to a fixed reference MR space via a learned field $\phi(\cdot,\cdot)$ and averaging random subsets of size $|S|$, where $|S|\in\{1,4,8,12\}$. These pseudo-PET images train a score-based diffusion model (SGM) as a subject-specific prior and are used in a diffusion-based reconstruction (PET-DDS) for single-subject PET data. Empirical results show visual improvements and reduced background noise in 2D reconstructions, surpassing OSEM, MAP-EM, and a state-of-the-art diffusion baseline, especially at larger $|S|$. The work highlights the potential of anatomically informed priors derived from MR data to enhance PET reconstructions and suggests future comparisons with explicit MR-guided reconstruction strategies.

Abstract

Large high-quality medical image datasets are difficult to acquire but necessary for many deep learning applications. For positron emission tomography (PET), reconstructed image quality is limited by inherent Poisson noise. We propose a novel method for synthesising diverse and realistic pseudo-PET images with improved signal-to-noise ratio. We also show how our pseudo-PET images may be exploited as a generative prior for single-subject PET image reconstruction. Firstly, we perform deep-learned deformable registration of multi-subject magnetic resonance (MR) images paired to multi-subject PET images. We then use the anatomically-learned deformation fields to transform multiple PET images to the same reference space, before averaging random subsets of the transformed multi-subject data to form a large number of varying pseudo-PET images. We observe that using MR information for registration imbues the resulting pseudo-PET images with improved anatomical detail compared to the originals. We consider applications to PET image reconstruction, by generating pseudo-PET images in the same space as the intended single-subject reconstruction and using them as training data for a diffusion model-based reconstruction method. We show visual improvement and reduced background noise in our 2D reconstructions as compared to OSEM, MAP-EM and an existing state-of-the-art diffusion model-based approach. Our method shows the potential for utilising highly subject-specific prior information within a generative reconstruction framework. Future work may compare the benefits of our approach to explicitly MR-guided reconstruction methodologies.

Figures (3)

• Figure 1: Outline of our reconstruction methodology: generating pseudo-PET images in our single-subject target space; incorporating target-specific pseudo-PET images as a prior into SGM-based reconstruction; and, reconstructing from target PET data.
• Figure 2: Gray-to-white matter contrast ratio plotted against normalised standard deviation in a background region of white matter, for iteration number 20 (unfilled) to 100 (filled).
• Figure 3: One transverse brain slice reconstructed with different variants of our methodology, with baseline images for comparison. OSEM 3D is reconstructed from fully 3D PET sinograms, while other PET images are reconstructed from a single 2D direct sinogram.