Abstract: Diffusion magnetic resonance imaging (dMRI) is widely used to study the microstructure and fiber tract orientation of white matter in the brain. High angular and multi-shell sampling with high spatial resolution typically requires longer scan times. In recent years, deep learning techniques have been extensively applied to dMRI super-resolution reconstruction, aiming to reconstruct high-resolution imaging signals from rapidly acquired images under sparse sampling conditions, thereby enabling more accurate fitting of brain microstructure imaging parameters. This paper surveys and analyzes the latest research progress in deep learning-based reconstruction of brain diffusion magnetic resonance images. According to different reconstruction targets, the methods are categorized into three types: reconstruction of basic diffusion metrics, reconstruction of high-order microstructure metrics, and reconstruction of the fiber orientation distribution function. For each category, the implementation techniques, evaluation metrics, and commonly used public datasets are discussed in detail. Finally, the main challenges and research trends in dMRI super-resolution reconstruction are summarized.

Key words: diffusion magnetic resonance imaging, deep learning, fiber orientation distribution function, microstructural metrics, super resolution