Supervise-assisted Multi-modality Fusion Diffusion Model for PET Restoration

TL;DR

This work addresses the challenge of restoring high-quality standard-dose PET from low-dose PET inputs by leveraging MR-guided multimodal information. It introduces MFdiff, a diffusion-based framework with a dedicated multi-modality feature fusion module and a conditional restoration process, enhanced by a two-stage supervise-assisted learning strategy that first learns generalized priors from simulated data and then tailors to in-vivo data. The method demonstrates superior performance over six baselines on simulated phantom data and in-vivo OOD datasets, with ablations confirming the importance of both global and detail fusion and the diffusion-based generator. The approach advances robustness to acquisition variations and offers practical potential for safer PET imaging, though further multi-center validation and pathological assessments are needed.

Abstract

Positron emission tomography (PET) offers powerful functional imaging but involves radiation exposure. Efforts to reduce this exposure by lowering the radiotracer dose or scan time can degrade image quality. While using magnetic resonance (MR) images with clearer anatomical information to restore standard-dose PET (SPET) from low-dose PET (LPET) is a promising approach, it faces challenges with the inconsistencies in the structure and texture of multi-modality fusion, as well as the mismatch in out-of-distribution (OOD) data. In this paper, we propose a supervise-assisted multi-modality fusion diffusion model (MFdiff) for addressing these challenges for high-quality PET restoration. Firstly, to fully utilize auxiliary MR images without introducing extraneous details in the restored image, a multi-modality feature fusion module is designed to learn an optimized fusion feature. Secondly, using the fusion feature as an additional condition, high-quality SPET images are iteratively generated based on the diffusion model. Furthermore, we introduce a two-stage supervise-assisted learning strategy that harnesses both generalized priors from simulated in-distribution datasets and specific priors tailored to in-vivo OOD data. Experiments demonstrate that the proposed MFdiff effectively restores high-quality SPET images from multi-modality inputs and outperforms state-of-the-art methods both qualitatively and quantitatively.

Figures (6)

• Figure 1: (a-b) Two examples of T1-MRI, LPET, and the corresponding SPET images from the phantom and in-vivo dataset. (c) The categorization of existing PET restoration methods. (d) Our MFdiff restoration method.
• Figure 2: Fig. 2 Overview of our supervise-assisted multi-modality fusion diffusion model (MFdiff). (a) provides the framework of MFdiff, which includes two main components: a conditional diffusion restoration module and a multi-modality feature fusion module. The details of the latter module are illustrated in (b) intra-modality learning module with modality ecoder and global/detailed encoder, and (c) cross-modality aggregation module.
• Figure 3: Visual comparison results of the restored PET along with error maps by different methods on phantom dataset.
• Figure 4: Visual comparison along with error maps of different methods on OOD data-1 with variation in imaging time.
• Figure 5: Visual comparison along with error maps of different methods on OOD data-2 with variation in radiotracer injection dose.