Deep diffusion prior for unsupervised super-resolution of diffusion-weighted images

Abstract

Diffusion-weighted imaging suffers from low resolution, which adversely affects subsequent quantitative analysis. Recent deep learning-based super-resolution methods have shown promising performance in enhancing the resolution of diffusion-weighted images. However, most existing methods rely on supervised learning, which requires paired high- and low-resolution images. Such images are often challenging to obtain in practice, greatly limiting the applicability of these methods. To overcome this limitation, we propose an innovative unsupervised diffusion-weighted image super-resolution method, called deep diffusion prior, which leverages intrinsic low-level image features to enhance the resolution of diffusion-weighted images using only low-resolution images. Our method integrates two key guiding priors to improve super-resolution performance: (1) Structural priors, derived from structural magnetic resonance images and track density images, to provide detailed anatomical guidance, and (2) Angular priors, modeled through angular neighboring constraints in the diffusion wavevector space, to provide effective angular guidance. Extensive experiments across diverse datasets, including the Human Connectome Project, the developing Human Connectome Project, and the Alzheimer’s Disease Neuroimaging Initiative, demonstrate that our method consistently outperforms existing methods in both qualitative and quantitative evaluations. These results highlight the potential of our method to advance more reliable quantitative analysis of high-quality resolution-enhanced diffusion-weighted images.