A Comprehensive Survey on Magnetic Resonance Image Reconstruction

TL;DR

This survey addresses the challenge of reconstructing high-quality MRI images from undersampled data, a problem with significant clinical impact. It surveys traditional reconstruction methods and a broad spectrum of deep learning–based approaches, including CNNs, U-Nets, GANs, Transformers, diffusion models, deep unfolding networks, and Mamba, as well as single- and multi-modal reconstruction, training strategies, and evaluation metrics. The article also reviews public datasets, downstream-task considerations, and practical challenges such as domain shift, data scarcity, computational cost, clinical adoption, interpretability, and generalization, offering a forward-looking discussion of potential directions. Key recommendations include exploring learnable undersampling masks, unsupervised and semi-supervised learning, multi-task and federated frameworks, and diffusion priors to better align reconstruction with real-world clinical workflows.

Abstract

Magnetic resonance imaging (MRI) reconstruction is a fundamental task aimed at recovering high-quality images from undersampled or low-quality MRI data. This process enhances diagnostic accuracy and optimizes clinical applications. In recent years, deep learning-based MRI reconstruction has made significant progress. Advancements include single-modality feature extraction using different network architectures, the integration of multimodal information, and the adoption of unsupervised or semi-supervised learning strategies. However, despite extensive research, MRI reconstruction remains a challenging problem that has yet to be fully resolved. This survey provides a systematic review of MRI reconstruction methods, covering key aspects such as data acquisition and preprocessing, publicly available datasets, single and multi-modal reconstruction models, training strategies, and evaluation metrics based on image reconstruction and downstream tasks. Additionally, we analyze the major challenges in this field and explore potential future directions.

Figures (13)

• Figure 1: This survey paper presents an analysis of the articles under investigation. The left section features a bar chart depicting the distribution of paper counts across different years; it is evident that the majority of the reviewed papers were introduced over the past five years. The right section provides a visual representation of the origins of the examined articles, highlighting that our survey encompasses a range of sources pertinent to the domain of MRI reconstruction.
• Figure 2: An overview of the paper organization.
• Figure 3: Some examples of undersampling masks. (a) Cartesian. (b) Gaussian. (c) Radial.
• Figure 4: Some examples of MRI images.
• Figure 5: The architecture diagram of RNLFNet zhou2023rnlfnet, a typical U-Net reconstruction network.