Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CRUNet-MR-Univ: A Foundation Model for Diverse Cardiac MRI Reconstruction

Donghang Lyu, Marius Staring, Hildo Lamb, Mariya Doneva

TL;DR

The paper addresses the challenge of generalizing cardiac MRI reconstruction across diverse scans, contrasts, trajectories, and centers. It introduces CRUNet-MR-Univ, a foundation-style, spatio-temporal unrolled model that fuses a bidirectional CRUNet with Cascaded Feature Aggregation and prompt-based priors to adapt to varied CMR data. Two-stage curriculum training on the CMRxRecon2025 dataset demonstrates improved generalization and center-wise performance over strong baselines, highlighting the value of temporal modeling and priors in robust reconstruction. This work advances practical robustness for clinical CMR reconstruction and points to future enhancements in receptive field depth and k-space-aware loss formulations.

Abstract

In recent years, deep learning has attracted increasing attention in the field of Cardiac MRI (CMR) reconstruction due to its superior performance over traditional methods, particularly in handling higher acceleration factors, highlighting its potential for real-world clinical applications. However, current deep learning methods remain limited in generalizability. CMR scans exhibit wide variability in image contrast, sampling patterns, scanner vendors, anatomical structures, and disease types. Most existing models are designed to handle only a single or narrow subset of these variations, leading to performance degradation when faced with distribution shifts. Therefore, it is beneficial to develop a unified model capable of generalizing across diverse CMR scenarios. To this end, we propose CRUNet-MR-Univ, a foundation model that leverages spatio-temporal correlations and prompt-based priors to effectively handle the full diversity of CMR scans. Our approach consistently outperforms baseline methods across a wide range of settings, highlighting its effectiveness and promise.

CRUNet-MR-Univ: A Foundation Model for Diverse Cardiac MRI Reconstruction

TL;DR

The paper addresses the challenge of generalizing cardiac MRI reconstruction across diverse scans, contrasts, trajectories, and centers. It introduces CRUNet-MR-Univ, a foundation-style, spatio-temporal unrolled model that fuses a bidirectional CRUNet with Cascaded Feature Aggregation and prompt-based priors to adapt to varied CMR data. Two-stage curriculum training on the CMRxRecon2025 dataset demonstrates improved generalization and center-wise performance over strong baselines, highlighting the value of temporal modeling and priors in robust reconstruction. This work advances practical robustness for clinical CMR reconstruction and points to future enhancements in receptive field depth and k-space-aware loss formulations.

Abstract

In recent years, deep learning has attracted increasing attention in the field of Cardiac MRI (CMR) reconstruction due to its superior performance over traditional methods, particularly in handling higher acceleration factors, highlighting its potential for real-world clinical applications. However, current deep learning methods remain limited in generalizability. CMR scans exhibit wide variability in image contrast, sampling patterns, scanner vendors, anatomical structures, and disease types. Most existing models are designed to handle only a single or narrow subset of these variations, leading to performance degradation when faced with distribution shifts. Therefore, it is beneficial to develop a unified model capable of generalizing across diverse CMR scenarios. To this end, we propose CRUNet-MR-Univ, a foundation model that leverages spatio-temporal correlations and prompt-based priors to effectively handle the full diversity of CMR scans. Our approach consistently outperforms baseline methods across a wide range of settings, highlighting its effectiveness and promise.
Paper Structure (12 sections, 2 equations, 2 figures, 3 tables)

This paper contains 12 sections, 2 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. CRUNet-MR-Univ
  4. Loss Function
  5. Experimental Setup
  6. Dataset and Task Description
  7. Implementation Details
  8. The First Stage
  9. The Second Stage
  10. Results
  11. Discussion and Conclusion
  12. Updated Experiment

Figures (2)

  • Figure 1: Overview of the CRUNet-MR-Univ model. The bottom section details the structure of each cascade block and its core components. Each cascade block contains a CRUNet model, with pink, gray, and green long-dashed lines showing the flow of hidden state features within the CRNNTI block at each level. Blue dotted and dashed lines represent text prompt inputs, while blue and red solid lines denote the flow of output undersampling-specific and spatial-specific prompt embeddings, respectively. The black dashed lines correspond to the input of the estimated CSM.
  • Figure 2: Visualizations of CRUNet-MR-Univ (proposed, S2) reconstruction results for three contrasts under three k-space trajectories at an acceleration factor of 24. Here, these cases are from validation set and lack ground truth references. 'REC' indicates the reconstructed images, and 'UND' is the original undersampled inputs.