HierAdaptMR: Cross-Center Cardiac MRI Reconstruction with Hierarchical Feature Adapters

Background: Clinical deployment of deep learning cardiac MRI reconstruction models encounters substantial performance degradation when applied across different medical centers due to heterogeneous scanner configurations, varying field strengths, and diverse imaging protocols. Existing domain adaptation strategies necessitate complete model retraining or compromise diagnostic fidelity through post-processing harmonization.

Methods: We introduce HierAdaptMR, addressing cross-center domain shifts through a dual-level adaptation architecture built on parameter-efficient adapters. Protocol-Level Adapters handle sequence-dependent variations across cine, LGE, mapping, and perfusion acquisitions, while Center-Level Adapters accommodate scanner-specific characteristics including field strength disparities (1.5T-5.0T), vendor implementations (GE, Philips, Siemens, UIH), and coil array configurations. The framework integrates HyperCMR's variational unrolling with PromptMR-plus enhancements, incorporating historical feature fusion to mitigate cascade information loss. A Universal Adapter facilitates zero-shot generalization to unseen centers through mixed training strategies. Multi-scale SSIM optimization with frequency domain weighting and contrast-adaptive loss functions ensures robust performance across imaging modalities.

Results: Evaluation across the CMRxRecon2025 multi-center dataset (5+ centers, 10+ scanners, 9 modalities) reveals consistent reconstruction improvements. Mean SSIM increased 12.9% with corresponding 13.0% PSNR gains compared to baseline methods. Performance scaling adapts to center characteristics—sites with challenging baseline metrics achieved up to 25.6% SSIM improvements, while high-performing centers maintained steady enhancement. The approach demonstrates vendor-agnostic effectiveness across all tested platforms without manufacturer-specific tuning. Parameter overhead remains minimal at 2.1M additional weights (3.2% increase), achieving substantial efficiency gains over conventional fine-tuning approaches.

Conclusion: HierAdaptMR establishes a practical framework for clinical MRI reconstruction deployment across heterogeneous medical environments. The hierarchical adaptation strategy preserves diagnostic quality while eliminating site-specific optimization requirements, addressing critical barriers to widespread adoption of deep learning reconstruction technologies in cardiac imaging.

Quantitative cross-center performance evaluation on CMRxRecon2025 dataset.

Center	Scanner Config	HyperCMR		HierAdaptMR		Improvement
		SSIM	PSNR	SSIM	PSNR	SSIM/PSNR
C001	UIH-3.0T	0.695	28.30	0.873	31.79	+25.6%/+12.3%
C002	Siemens-3.0T	0.682	25.01	0.797	28.34	+16.9%/+13.3%
	UIH-3.0T	0.770	27.44	0.856	31.20	+11.2%/+13.7%
C003	UIH-3.0T	0.808	29.31	0.878	32.86	+8.7%/+12.1%
C004	Siemens-1.5T	0.726	26.38	0.829	29.97	+14.2%/+13.6%
C005	GE-1.5T	0.846	30.30	0.904	33.78	+6.9%/+11.5%
	Siemens-3.0T	0.745	27.14	0.874	31.18	+17.3%/+14.9%
C006	Siemens-3.0T	0.753	27.40	0.866	31.14	+15.0%/+13.6%
	UIH-3.0T	0.845	29.62	0.916	33.65	+8.4%/+13.6%
C008	GE-1.5T	0.823	28.83	0.886	32.12	+7.7%/+11.4%
Overall Mean		0.769	27.97	0.868	31.60	+12.9%/+13.0%

(a) Hierarchical Feature Adaptation Framework. The architecture integrates HyperCMR's variational unrolling with dual-level adaptation: Protocol-Level Adapters address sequence-dependent variations while Center-Level Adapters handle scanner-specific differences. A Universal Adapter enables generalization to entirely unseen centers during evaluation. (b) (c) Historical feature fusion architecture for enhanced cascade-based reconstruction. The proposed framework addresses feature information loss in traditional cascade networks through adaptive attention mechanisms and multi-scale memory integration.
Qualitative cross-center reconstruction results on CMRxRecon2025 validation set. Each column represents a different center configuration with varying undersampling patterns. From top to bottom: masked k-space with undersampling artifacts, recovered k-space, and final reconstructed cardiac images.