Generalizable Unpaired PET Image Enhancement via Two-Stage Diffusion Framework

Positron emission tomography (PET) image enhancement is important for improving clinical diagnostic accuracy. However, most existing enhancement methods rely on paired datasets, which are difficult to acquire in clinical practice. This study aims to develop a robust unpaired framework for PET image enhancement that can effectively leverage the widely available unpaired low- and high-quality datasets. By eliminating the need for paired data, the proposed framework can generalize across diverse scanners, protocols, and radiotracer doses. We propose a two-stage diffusion-based framework for unpaired PET image enhancement. In Stage 1, an unconditional diffusion model is trained on low-quality data to learn their distribution. During sampling, unpaired high-quality data guides the generation of pseudo low-quality counterparts via a conditional strategy combining timestep truncation and frequency-domain constraints, constructing a synthetic paired dataset. In Stage 2, a denoising network (e.g., RED-CNN) is trained on this synthetic dataset for efficient enhancement. The proposed method consistently outperforms traditional approaches (BM3D, BM4D), unsupervised diffusion-based methods (ILVR, SDEdit, EGSDE), self-supervised denoising methods (Noise2Void, Conditional Deep Image Prior(CDIP)), and unpaired methods (IE-CycleGAN) on both low-quality brain datasets from the uMI 780 scanner (comprising 420 subjects) and the Biograph mMR scanner (comprising 19 subjects). Compared with supervised methods, our approach demonstrates superior generalization ability and exhibits better applicability to real-world clinical datasets lacking paired training data. The proposed method effectively harnesses the powerful generative capacity of diffusion models while bypassing the time-consuming inference issue inherent to direct application of diffusion models through a flexible second stage. This provides a robust and scalable framework for enhancing real-world low-quality PET images without requiring paired data.