Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Steerable Conditional Diffusion for Domain Adaptation in PET Image Reconstruction

George Webber, Alexander Hammers, Andrew P. King, Andrew J. Reader

TL;DR

This work addresses domain shift in diffusion-based PET reconstruction by integrating steerable conditional diffusion (SCD) with likelihood-scheduled diffusion (PET-LiSch) to enable per-scan prior adaptation without retraining. At reconstruction time, a low-rank adaptation (LoRA) module updates the diffusion prior per scan while maintaining data fidelity via a Poisson likelihood model and a likelihood schedule, with end-point alignment refined through Tweedie’s formula. In synthetic 2D brain phantoms with a structured domain shift, the proposed PET-LiSch-SCD method suppresses artefacts common to unsteered diffusion methods and outperforms MLEM and PET-LiSch at matched likelihood, demonstrating robustness to domain shift. The approach offers a practical path toward clinically robust diffusion-based PET reconstruction and motivates evaluation on 3D and real patient data.

Abstract

Diffusion models have recently enabled state-of-the-art reconstruction of positron emission tomography (PET) images while requiring only image training data. However, domain shift remains a key concern for clinical adoption: priors trained on images from one anatomy, acquisition protocol or pathology may produce artefacts on out-of-distribution data. We propose integrating steerable conditional diffusion (SCD) with our previously-introduced likelihood-scheduled diffusion (PET-LiSch) framework to improve the alignment of the diffusion model's prior to the target subject. At reconstruction time, for each diffusion step, we use low-rank adaptation (LoRA) to align the diffusion model prior with the target domain on the fly. Experiments on realistic synthetic 2D brain phantoms demonstrate that our approach suppresses hallucinated artefacts under domain shift, i.e. when our diffusion model is trained on perturbed images and tested on normal anatomy, our approach suppresses the hallucinated structure, outperforming both OSEM and diffusion model baselines qualitatively and quantitatively. These results provide a proof-of-concept that steerable priors can mitigate domain shift in diffusion-based PET reconstruction and motivate future evaluation on real data.

Steerable Conditional Diffusion for Domain Adaptation in PET Image Reconstruction

TL;DR

This work addresses domain shift in diffusion-based PET reconstruction by integrating steerable conditional diffusion (SCD) with likelihood-scheduled diffusion (PET-LiSch) to enable per-scan prior adaptation without retraining. At reconstruction time, a low-rank adaptation (LoRA) module updates the diffusion prior per scan while maintaining data fidelity via a Poisson likelihood model and a likelihood schedule, with end-point alignment refined through Tweedie’s formula. In synthetic 2D brain phantoms with a structured domain shift, the proposed PET-LiSch-SCD method suppresses artefacts common to unsteered diffusion methods and outperforms MLEM and PET-LiSch at matched likelihood, demonstrating robustness to domain shift. The approach offers a practical path toward clinically robust diffusion-based PET reconstruction and motivates evaluation on 3D and real patient data.

Abstract

Diffusion models have recently enabled state-of-the-art reconstruction of positron emission tomography (PET) images while requiring only image training data. However, domain shift remains a key concern for clinical adoption: priors trained on images from one anatomy, acquisition protocol or pathology may produce artefacts on out-of-distribution data. We propose integrating steerable conditional diffusion (SCD) with our previously-introduced likelihood-scheduled diffusion (PET-LiSch) framework to improve the alignment of the diffusion model's prior to the target subject. At reconstruction time, for each diffusion step, we use low-rank adaptation (LoRA) to align the diffusion model prior with the target domain on the fly. Experiments on realistic synthetic 2D brain phantoms demonstrate that our approach suppresses hallucinated artefacts under domain shift, i.e. when our diffusion model is trained on perturbed images and tested on normal anatomy, our approach suppresses the hallucinated structure, outperforming both OSEM and diffusion model baselines qualitatively and quantitatively. These results provide a proof-of-concept that steerable priors can mitigate domain shift in diffusion-based PET reconstruction and motivate future evaluation on real data.
Paper Structure (5 sections, 1 figure, 1 algorithm)

This paper contains 5 sections, 1 figure, 1 algorithm.

Table of Contents

  1. Introduction
  2. Theory
  3. Experimental setup
  4. Results
  5. Summary

Figures (1)

  • Figure 1: Qualitative comparison on low-count reconstructions, with all reconstructions at the same likelihood corresponding to the $40^{\text{th}}$ iteration of MLEM. Column 1: the checkerboard artefact added to training images, shown applied to the ground truth. Column 2: ground truth image. Column 3: image reconstructed by MLEM (40 its). Column 4: image reconstructed with likelihood-scheduled diffusion (likelihood matched to MLEM image). Note the presence of the checkerboard artefact along the right-hand side of the image and elsewhere. Column 5: proposed method PET-LiSch-SCD, without checkerboard artefact.