Can Natural Domain Foundation Models Be Applied to Cardiac MRI Reconstruction?

Anam Hashmi, Julia Dietlmeier, Mayug Maniparambil, Carles Garcia-Cabrera, Kathleen M Curran, Noel O'Connor

Keywords: Cardiac MRI Reconstruction, Foundation Models, U-Net

TL;DR: This paper explores the use of frozen vision foundation models, such as CLIP and DINOv2, for cardiac MRI reconstruction, demonstrating their potential to outperform traditional methods while highlighting areas for further improvement.

Abstract:

The field of computer vision has experienced a paradigm shift with the emergence of general-
purpose foundation models, which exhibit strong generalization capabilities across a wide
range of tasks. However, their applicability to specialized medical imaging tasks, particularly cardiac MRI reconstruction, remains underexplored. In this work, we investigate
the transferability of state-of-the-art vision foundation models like CLIP and DINOv2 for
cardiac MRI reconstruction. We propose a novel framework that leverages frozen vision
foundation models as image encoders, combined with a UNETR-based trainable decoder.
We validate our framework on the CMRxRecon2024 dataset, demonstrating improved performance over the traditional state-of-the-art U-Net under acceleration factor (×4), despite
relying on frozen natural-domain foundation model and significantly fewer trainable parameters. The code will be released at https://github.com/Hashmi360/MRI_Recon_Foundation_Models

Submission Number: 83