UEPS: Robust and Efficient MRI Reconstruction

Xiang Zhou; Hong Shang; Zijian Zhan; Tianyu He; Jintao Meng; Dong Liang

eess.IVMar 19, 2026

UEPS: Robust and Efficient MRI Reconstruction

Authors: Xiang Zhou, Hong Shang, Zijian Zhan, Tianyu He, Jintao Meng, Dong Liang

Organizations: Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China · School of Biomedical Engineering, Shenzhen University Medical School, Shenzhen University, China · University of Chinese Academy of Sciences, China · Faculty of Computer Science and Artificial Intelligence, Shenzhen University of Advanced Technology, China · State Key Laboratory of Biomedical Imaging Science and System, Chinese Academy of Sciences, China

Abstract

Deep unrolled models (DUM) have become the state-of-the-art for accelerated MRI reconstruction, yet their robustness under domain shift remains a critical barrier to clinical adoption. In this work, we identify coil sensitivity map (CSM) estimation as the primary bottleneck limiting generalization. To address this, we propose UEPS, a novel DUM architecture featuring three key designs: (i) an Unrolled Expanded (UE) design that eliminates CSM dependency by reconstructing each coil independently; (ii) progressive resolution, which leverages k-space-to-image mapping for efficient coarse-to-fine refinement; and (iii) sparse attention tailored to MRI's 1D undersampling nature. These physics-grounded designs enable gains in robustness and computational efficiency. We construct a large-scale zero-shot transfer benchmark comprising 10 out-of-distribution (OOD) test sets spanning diverse clinical shifts, anatomy, view, contrast, vendor, field strength, and coil configurations. Extensive experiments demonstrate that UEPS consistently and substantially outperforms existing DUM, end-to-end, diffusion, and untrained methods across all OOD tests, achieving state-of-the-art robustness with low-latency inference suitable for real-time deployment.

Figures & tables

Figure 1: Zero-shot transfer benchmark results. Average PSNR: mean over 10 OOD sets. Runtime: measured on FastMRI Knee (15 coils). Compared methods include E2EVarNet (MICCAI 2020), PDAC (CVPR 2024), NoVar (CVPR 2025), DDS (ICLR 2024), AeSPa (CVPR 2025), and SCAMPI (MRM 2024), alongside UNet and ViT baselines.

Figure 2: Overview of the proposed UEPS architecture, illustrating progressive resolution across cascades (e.g., 64 × 64, 128 × 128, 256 × 256 and 320 × 320) (a); schematic of measurement-informed k-space padding (b); and comparison of attention patterns, with different colours indicating the attention regions of different query tokens (c).

Figure 3: Attention maps from UE with a full-attention ViT backbone.

Figure 7: Reconstruction examples from the zero-shot transfer benchmark, each with PSNR/SSIM shown in the top-left corner.

Table 1: Ablation of progressive resolution and sparse attention. The final row corresponds to UEPS. Best values are bold and second best underlined.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Split	Dataset Name	Part	View	Sequence	Vendor	B0	Coils	Subj.	Slices
Train	FastMRI Brain	brain	axial	2D FSE	Siemens	3T/1.5T	2–28	4469	70748
Test(ID)	FastMRI Brain	brain	axial	2D FSE	Siemens	3T/1.5T	2–24	558	8852
Test (OOD)	FastMRI Knee	knee	coronal	2D FSE	Siemens	3T/1.5T	15	150	5427
	Stanford 2D	mix	mix	2D FSE	GE	3T	3–32	89	2037
	CC-359(Axi)	brain	axial	3D MP-RAGE	GE	3T	12	67	13802
	CC-359(Sag)	brain	sagittal	3D MP-RAGE	GE	3T	12	67	9490

Appendix

Table 4: Summary of MRI datasets. 3D datasets are treated as multi-slice 2D data in selected anatomical views: CC-359 is converted to axial and sagittal views as two separate test sets, and AHEAD to the axial view. The AHEAD sequence contains five echoes, each forming a separate test set. Sources: FastMRI ( Zbontar et al., 2018 ) , Stanford 2D ( Cheng, 2018 ) , CC-359 ( Souza et al., 2018 ) , M4Raw ( Lyu et al., 2023 ) , and AHEAD ( Alkemade et al., 2020 ) .

		CSM source
Dataset	Metric	BART	E2EVarNet	PDAC	NoVar
FastMRI Brain [ID]	PSNR $\uparrow$	44.95 $\pm 6.52$	44.28 $\pm 4.59$	43.84 $\pm 4.67$	41.98 $\pm 4.33$
FastMRI Brain [ID]	SSIM $\uparrow$	99.00 $\pm 2.03$	98.97 $\pm 2.30$	98.85 $\pm 2.58$	98.85 $\pm 2.48$
CC-359 Axi	PSNR $\uparrow$	41.86 $\pm 3.10$	38.96 $\pm 3.30$	37.22 $\pm 3.72$	38.83 $\pm 2.94$
CC-359 Axi	SSIM $\uparrow$	98.74 $\pm 0.96$	98.34 $\pm 1.02$	97.76 $\pm 1.62$	98.41 $\pm 0.92$
CC-359 Sag	PSNR $\uparrow$	42.02 $\pm 3.24$	38.05 $\pm 3.02$	37.63 $\pm 3.22$	36.89 $\pm 2.52$

Appendix

Table 5: Complete CSM evaluation results on the ID and ten OOD test sets. Values are mean

\pm

standard deviation across valid slices. Average is computed over the ten OOD test sets. PSNR is reported in dB and SSIM in %. Best values are bold and second best underlined.

Dataset	Metric	UNet	ViT	E2EVarNet	PDAC	NoVar	DDS	AeSPa	SCAMPI	UEPS(Ours)
		Model
FastMRI Brain [ID]	PSNR(dB) $\uparrow$	40.58 $\pm 3.02$	41.42 $\pm 2.81$	41.93 $\pm 2.86$	41.03 $\pm 2.81$	40.70 $\pm 2.87$	35.91 $\pm 5.37$	35.68 $\pm 2.61$	36.48 $\pm 2.42$	41.87 $\pm 2.87$
FastMRI Brain [ID]	SSIM(%) $\uparrow$	95.61 $\pm 3.49$	95.95 $\pm 3.49$	96.27 $\pm 3.34$	95.69 $\pm 3.63$	95.68 $\pm 3.62$	85.52 $\pm 21.49$	89.05 $\pm 5.16$	76.48 $\pm 10.54$	96.15 $\pm 3.49$
CC-359 Axi	PSNR(dB) $\uparrow$	29.09 $\pm 3.45$	30.99 $\pm 3.52$	31.27 $\pm 3.46$	31.73 $\pm 3.48$	30.65 $\pm 3.50$	29.17 $\pm 5.42$	30.11 $\pm 2.47$	29.83 $\pm 2.52$	32.72 $\pm 3.25$
CC-359 Axi	SSIM(%) $\uparrow$	83.40 $\pm 6.54$	87.34 $\pm 5.19$	87.86 $\pm 5.16$	87.36 $\pm 5.09$	86.03 $\pm 5.76$	74.10 $\pm 25.72$	82.16 $\pm 3.70$	73.17 $\pm 7.98$	90.35 $\pm 3.23$
CC-359 Sag	PSNR(dB) $\uparrow$	27.36 $\pm 5.08$	31.98 $\pm 3.32$	32.86 $\pm 3.32$	32.99 $\pm 3.30$	31.57 $\pm 3.47$	30.23 $\pm 4.72$	31.82 $\pm 2.13$	30.59 $\pm 2.96$	34.45 $\pm 3.33$

Appendix

Table 6: Complete results for the zero-shot transfer benchmark in fig. 1 . Average is computed over the ten OOD test sets. Best values are in bold and second best underlined.

Figure 9: Additional reconstruction examples, each with PSNR/SSIM shown in the top-left corner (Part 1).

Figure 10: Additional reconstruction examples, each with PSNR/SSIM shown in the top-left corner (Part 2).

Explore similar work

Deep Unrolled Networks in Representation Space Applied to MRI Reconstruction
Jun 19, 2026Efe Ilıcak, Baris Imre, Chloé Najac +4Accelerated MRI Reconstruction Image Inverse Problems
Department of Radiology, Leiden University Medical Center, Leiden, Netherlands
UMPIRE-Net: Unrolled Magnitude-Phase Regularization Network for Accelerated MRI
Aug 14, 2026Mahdi Saberi, Toygan Kiliç, Mehmet AkçakayaAccelerated MRI Reconstruction Physics-Informed ML
Department of Electrical and Computer Engineering, University of Minnesota, Minneapolis, MN, USA · Center for Magnetic Resonance Research, University of Minnesota, Minneapolis, MN, USA
Robustness of breast lesion segmentation under MRI undersampling improves with k-space-aware deep learning
May 21, 2026Lukas T. Rotkopf, Marco Schlimbach, Julius C. Holzschuh +3Magnetic Resonance Imaging Medical Image Segmentation
Institute for AI in Medicine (IKIM), University Hospital Essen, Essen, Germany · Department of Physics, Technical University Dortmund, Dortmund, Germany · 3Division of Radiology, German Cancer Research Center (DKFZ), Heidelberg, Germany +4

Jun 19, 2026J/K move · Enter open · S save

Deep Unrolled Networks in Representation Space Applied to MRI Reconstruction

Efe Ilıcak, Baris Imre, Chloé Najac, Ruben van den Broek+3

Department of Radiology, Leiden University Medical Center, Leiden, Netherlands

Accelerated MRI Reconstruction Image Inverse Problems

Open paper AskPDF arXiv

Deep unrolled networks (DUNs) integrate physical forward models with learned regularization in cascaded network architectures, achieving exceptional performance in inverse problems while maintaining interpretability. While most DUNs operate in the object domain (e.g., image space), recent variants explored representation spaces for improved information flow. However, these methods rely on heuristic methods for data consistency (DC), sacrificing fidelity with measurements. In this work, we introduce DUNE (Deep Unrolled Networks in rEpresentation space), a framework that maintains exact adherence to physical measurements while operating in learned representation spaces. By deriving the DC gradient via the chain rule and implementing it through the Vector-Jacobian Product (VJP), we enable exact backpropagation of measurement residuals into the representation space. This formulation supports diverse architectural backbones, including pre-trained encoders to guide the iterative process. We assess DUNE against state-of-the-art baselines on accelerated MRI reconstruction tasks, demonstrating that exact VJP-based gradients yield superior reconstruction quality and structural fidelity across both single-channel portable low-field and multi-channel clinical high-field MRI acquisitions. The code will be available upon publication at https://github.com/EfeIlicak/DUNE.

Abstract

Explore similar work

Deep Unrolled Networks in Representation Space Applied to MRI Reconstruction

UMPIRE-Net: Unrolled Magnitude-Phase Regularization Network for Accelerated MRI

Robustness of breast lesion segmentation under MRI undersampling improves with k-space-aware deep learning