Foundation models improve data-efficient cardiac MRI image reconstruction
Beyond Training from Scratch: Foundation Models for Data-Efficient and Generalizable Cardiac MRI Reconstruction
Computer Vision and Pattern Recognition
Summary
Cardiac MRI scans can be sped up by reconstructing images from partial data, but this usually needs lots of training data and struggles when data is limited or different from training. The authors show that using pretrained vision models, originally built for other image tasks, helps reconstruct good cardiac MRI images with less data and better adaptation to new settings. They found that models like DINOv2 perform best and that different ways of adjusting these models help depending on how much data is available. This approach could make faster and more reliable heart MRI scans more accessible.
What this means in practice
- •For medical imaging teams: Improve cardiac MRI reconstruction quality and speed when training data is limited or comes from different hospitals using pretrained vision models.
- •For radiology software developers: Integrate foundation models like DINOv2 into MRI reconstruction software to enable faster heart scans with better image fidelity under varied conditions.$Commercial implications: This paper enables development of commercial MRI reconstruction tools that reduce scan times while maintaining image quality, benefiting clinical workflows.
Authors
Anam Hashmi, Mayug Maniparambil, Julia Dietlmeier, Kathleen M. Curran, Noel E. O'Connor
Abstract
Cardiac magnetic resonance imaging reconstruction aims to recover high-quality images from undersampled acquisitions, enabling faster scans while preserving diagnostic fidelity. Recent reconstruction methods are typically trained from scratch and often require large amounts of task-specific data, limiting their robustness under data scarcity and distribution shifts. In this work, we investigate whether pretrained vision foundation models can serve as effective priors for accelerated cardiac MRI reconstruction. We propose a reconstruction framework that integrates frozen and parameter-efficiently adapted visual encoders, including CLIP, BiomedCLIP, and DINOv2, within a transformer-based reconstruction architecture. Extensive experiments on the CMRxRecon2023 and CMRxRecon2024 benchmarks demonstrate that pretrained representations consistently outperform a transformer trained from scratch across multiple acceleration factors. We further evaluate performance under limited supervision and cross-dataset transfer, showing that foundation models provide superior data efficiency and generalization. While frozen representations are particularly effective in extreme low-data regimes, Low-Rank Adaptation (LoRA) yields additional gains when moderate amounts of training data are available. Among the evaluated backbones, DINOv2 achieves the strongest overall performance. These findings highlight the potential of vision foundation models as robust and transferable priors for cardiac MRI reconstruction.