Beyond Training from Scratch: Foundation Models for Data-Efficient and Generalizable Cardiac MRI Reconstruction

Anam Hashmi; Mayug Maniparambil; Julia Dietlmeier; Kathleen M. Curran; Noel E. O'Connor

cs.CVOct 6, 2026

Beyond Training from Scratch: Foundation Models for Data-Efficient and Generalizable Cardiac MRI Reconstruction

Authors: Anam Hashmi, Mayug Maniparambil, Julia Dietlmeier, Kathleen M. Curran, Noel E. O'Connor

Organizations: Research Ireland Centre for Research Training in Machine Learning, Ireland · Dublin City University, Ireland · University College Dublin, Ireland · Rinn Artificial Intelligence, Ireland

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.

Figures & tables

Figure 1 : Overview of the baseline UNETR architecture (left) and the proposed foundation model reconstruction framework (right). The scratch-trained transformer encoder of UNETR is replaced with a pretrained vision foundation model (CLIP, BiomedCLIP, or DINOv2). Features from all transformer layers are fused to form the bottleneck representation, while intermediate features provide multi-scale skip connections to the decoder. The encoder is evaluated in both frozen and LoRA-adapted settings, enabling a controlled study of representation transfer for cardiac MRI reconstruction.

Acc.	Model	CMRxRecon2023			CMRxRecon2024
		SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$
4 $\times$	UNETR [65.7M]	0.8904	31.1405	0.0458	0.8801	30.7491	0.0552
	Frozen CLIP [33.7M]	0.9035	32.1443	0.0344	0.8950	31.4805	0.0448
	Frozen BiomedCLIP [33.7M]	0.9027	32.1010	0.0351	0.8911	31.1441	0.0495
	Frozen DINOv2 [33.7M]	0.9073	32.2097	0.0348	0.8979	31.5111	0.0456
8 $\times$	UNETR	0.8625	29.7299	0.0632	0.8412	29.0134	0.0831

Table 1 : Quantitative reconstruction performance on CMRxRecon2023 and CMRxRecon2024. Best results are shown in bold.

Acc.	Model	CMRxRecon2023			CMRxRecon2024
		SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$
4 $\times$	Frozen CLIP	0.9035	32.1443	0.0344	0.8950	31.4805	0.0448
	LoRA CLIP	0.9013 ↓	32.1458 ↑	0.0340 ↑	0.9003 ↑	32.2644 ↑	0.0377 ↑
	Frozen BiomedCLIP	0.9027	32.1010	0.0351	0.8911	31.1441	0.0495
	LoRA BiomedCLIP	0.8972 ↓	31.9133 ↓	0.0352 ↓	0.8945 ↑	31.8295 ↑	0.0407 ↑
	Frozen DINOv2	0.9073	32.2097	0.0348	0.8979	31.5111	0.0456

Table 2 : Effect of LoRA adaptation on pretrained foundation models across CMRxRecon2023 and CMRxRecon2024. Arrows indicate the change relative to the corresponding frozen model: green arrows (

\downarrow

) denote performance degradation, while red arrows (

\uparrow

) denote an improvement. Best results for each acceleration factor and dataset are shown in bold.

Data	Model	CMRxRecon2023			CMRxRecon2024
Data	Model	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$
5%	UNETR	0.7900	27.1943	0.1098	0.7803	26.6047	0.1445
	Frozen CLIP	0.8310	28.2532	0.0835	0.8116	27.8297	0.1007
	LoRA CLIP	0.8196	27.7251	0.0977	0.8061	27.7494	0.0991
	Frozen BiomedCLIP	0.8199	27.7550	0.0970	0.8056	27.5921	0.1055
	LoRA BiomedCLIP	0.8168	27.5929	0.1025	0.8011	26.8955	0.1346

Table 3 : Low-data regime analysis at

10\times

acceleration on CMRxRecon2023 and CMRxRecon2024. Results are reported using 5%, 25%, 50%, and 75% of the available training data. Best results within each data regime are highlighted in bold.

Figure 2 : Low-data regime analysis on CMRxRecon2024 at $10\times$ acceleration. Foundation model-based methods consistently outperform UNETR across all training fractions. Frozen DINOv2 achieves the strongest performance in the extreme 5% regime, while LoRA-adapted DINOv2 becomes increasingly advantageous as more training data becomes available.

Model	CMRxRecon2024 $\rightarrow$ CMRxRecon2023			CMRxRecon2023 $\rightarrow$ CMRxRecon2024
Model	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$
UNETR	0.8293	27.9917	0.0849	0.7596	25.8888	0.1833
Frozen CLIP	0.8510	28.8955	0.0686	0.7971	27.2573	0.1183
LoRA CLIP	0.8533	29.1395	0.0639	0.7919	27.3316	0.1117
Frozen BiomedCLIP	0.8360	28.2405	0.0796	0.7780	26.1926	0.1569
LoRA BiomedCLIP	0.8473	28.3863	0.0822	0.7889	27.2543	0.1155

Table 4 : Cross-dataset generalization results at

10\times

acceleration. Models are trained on one challenge dataset and evaluated on the other dataset. Best results for each transfer direction are shown in bold.

Figure 3 : Qualitative reconstruction results and absolute error maps. Left: In-domain performance on CMRxRecon2024 at $4\times$ acceleration, where foundation models produce sharper reconstructions and lower errors than UNETR. Right: Cross-dataset transfer from CMRxRecon2024 to CMRxRecon2023 at $10\times$ acceleration, showing improved robustness and anatomical fidelity for LoRA-adapted foundation models. Red boxes indicate representative regions of interest.

Dataset	Rank	Params (M)	$4\times$			$8\times$
Dataset	Rank	Params (M)	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$
CMRxRecon2023	8	34.0	0.8973	31.8867	0.0356	0.8679	30.3138	0.0508
	64	36.1	0.8975	31.8523	0.0375	0.8672	29.9981	0.0586
	128	38.4	0.8972	31.9133	0.0351	0.8706	30.4793	0.0516
CMRxRecon2024	8	34.0	0.8917	31.6805	0.0416	0.8484	29.5407	0.0707
	64	36.1	0.8925	31.8170	0.0409	0.8509	29.8524	0.0640

Table 5 : LoRA rank ablation for BiomedCLIP on CMRxRecon2023 and CMRxRecon2024. Trainable parameter counts are reported in millions (M). Trainable parameter counts include the LoRA modules, feature fusion layers, and reconstruction decoder. The best-performing rank for each acceleration factor is highlighted in bold.

Dataset	Rank	Params (M)	$4\times$			$8\times$
Dataset	Rank	Params (M)	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$
CMRxRecon2023	8	34.1	0.9020	32.0410	0.0351	0.8699	30.5080	0.0498
	64	37.2	0.9009	32.0592	0.0344	0.8703	30.2384	0.0530
	128	40.8	0.9013	32.1458	0.0340	0.8710	30.4831	0.0492
CMRxRecon2024	8	34.1	0.8967	31.9873	0.0394	0.8571	29.9309	0.0645
	64 (for $8\times$ )	37.2	0.8992	32.1980	0.0374	0.8591	30.2022	0.0585

Table 6 : LoRA rank ablation for CLIP on CMRxRecon2023 and CMRxRecon2024. Trainable parameter counts are reported in millions (M). Trainable parameters include the LoRA modules, feature fusion layers, and reconstruction decoder. The best-performing rank for each acceleration factor is highlighted in bold.

Dataset	Rank	Params (M)	$4\times$			$8\times$
Dataset	Rank	Params (M)	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	NMSE $\downarrow$
CMRxRecon2023	8	34.1	0.9050	32.4577	0.0316	0.8768	30.3741	0.0524
	64	37.2	0.9058	32.2943	0.0324	0.8758	30.7677	0.0452
	128	40.8	0.9061	32.4175	0.0319	0.8776	30.9313	0.0441
CMRxRecon2024	8	34.1	0.9023	32.2638	0.0365	0.8628	30.0431	0.0623
	64	37.2	0.9025	32.2664	0.0379	0.8636	30.0201	0.0632

Table 7 : LoRA rank ablation for DINOv2 on CMRxRecon2023 and CMRxRecon2024. Trainable parameter counts are reported in millions (M). Trainable parameters include the LoRA modules, feature fusion layers, and reconstruction decoder. The best-performing rank for each acceleration factor is highlighted in bold.

Explore similar work

Are Natural-Domain Foundation Models Effective for Accelerated Cardiac MRI Reconstruction?
Apr 24, 2026Anam Hashmi, Mayug Maniparambil, Julia Dietlmeier +2Domain Generalization Vision Foundation Model Adaptation
1Dublin City University, Ireland · University College Dublin, Ireland
CMRVision: A Foundation Model for Cardiac MR Image Analysis
Sep 1, 2026Athira J. Jacob, Puneet Sharma, Daniel RueckertMedical Imaging Foundation Models Medical Image Classification
Digital Technology and Innovation, Siemens Healthineers, Princeton, NJ, USA · Chair for Al in Healthcare and Medicine, Technical University of Munich (TUM) and TUM University Hospital, Munich, Germany · Department of Computing, Imperial College London, UK
Enhancing Implicit Neural Representations with Image Feature Embedding for Unsupervised Cardiac Cine MRI Reconstruction
Jul 5, 2026Donghang Lyu, Marius Staring, Yiming Dong +3Implicit Neural Representations Accelerated MRI Reconstruction
Division of Image Processing, Department of Radiology, Leiden University Medical Center, 2333ZA Leiden, Netherlands · C.J. Gorter Center, Department of Radiology, Leiden University Medical Center, 2333ZA Leiden, Netherlands · Philips Innovative Technologies, 22335 Hamburg, Germany +1

Apr 24, 2026, eess.IVJ/K move · Enter open · S save

Are Natural-Domain Foundation Models Effective for Accelerated Cardiac MRI Reconstruction?

Anam Hashmi, Mayug Maniparambil, Julia Dietlmeier, Kathleen M. Curran+1

1Dublin City University, Ireland · University College Dublin, Ireland

Domain Generalization Vision Foundation Model Adaptation Accelerated MRI Reconstruction Medical Image Representation Learning Cardiac MRI

Open paper AskPDF arXiv

The emergence of large-scale pretrained foundation models has transformed computer vision, enabling strong performance across diverse downstream tasks. However, their potential for physics-based inverse problems, such as accelerated cardiac MRI reconstruction, remains largely underexplored. In this work, we investigate whether natural-domain foundation models can serve as effective image priors for accelerated cardiac MRI reconstruction, and compare the performance obtained against domain-specific counterparts such as BiomedCLIP. We propose an unrolled reconstruction framework that incorporates pretrained, frozen visual encoders, such as CLIP, DINOv2, and BiomedCLIP, within each cascade to guide the reconstruction process. Through extensive experiments, we show that while task-specific state-of-the-art reconstruction models such as E2E-VarNet achieve superior performance in standard in-distribution settings, foundation-model-based approaches remain competitive. More importantly, in challenging cross-domain scenarios, where models are trained on cardiac MRI and evaluated on anatomically distinct knee and brain datasets--foundation models exhibit improved robustness, particularly under high acceleration factors and limited low-frequency sampling. We further observe that natural-image-pretrained models, such as CLIP, learn highly transferable structural representations, while domain-specific pretraining (BiomedCLIP) provides modest additional gains in more ill-posed regimes. Overall, our results suggest that pretrained foundation models offer a promising source of transferable priors, enabling improved robustness and generalization in accelerated MRI reconstruction.

Abstract

Explore similar work

Are Natural-Domain Foundation Models Effective for Accelerated Cardiac MRI Reconstruction?

CMRVision: A Foundation Model for Cardiac MR Image Analysis

Enhancing Implicit Neural Representations with Image Feature Embedding for Unsupervised Cardiac Cine MRI Reconstruction