PU-CDM: A pyramid UNet based conditional diffusion model for sparse-view reconstruction in EPRI

Sparse-view reconstruction in electron paramagnetic resonance imaging (EPRI) aims to reduce scanning times, which is critical for tumor oxygen imaging, yet is often plagued by streak artifacts in filtered back-projection (FBP) reconstructions. To address this, we propose a pyramid UNet based conditional diffusion model (PU-CDM) to suppress these streak artifacts in EPRI images. PU-CDM uniquely introduces pyramid pooling and aggregation into the UNet architecture of the conditional diffusion model, while incorporating two advanced mechanisms—dense convolutions and self-attention—into the input module. By significantly improving the accuracy of the Gaussian noise prediction network in the conditional diffusion model, PU-CDM achieves superior performance in sparse-view reconstruction, generating high-quality images with only 5 sampling steps. Experimental results, both qualitative and quantitative, show that the images reconstructed by PU-CDM outperform those reconstructed by some existing representative deep learning models in terms of artifact removal and structural fidelity. PU-CDM can achieve accurate sparse-view reconstruction in EPRI, thus promoting EPRI towards fast scanning. In addition, PU-CDM can also be used for fast magnetic resonance imaging (MRI), low-dose computed tomography (LDCT) reconstruction, and natural image processing.