Organizations: All authors are with Cedars-Sinai Medical Center, Los Angeles, CA 90048 USA. Yimeng He is with the Biomedical Imaging Research Institute and the Department of Computational Biomedicine · All authors are with Cedars-Sinai Medical Center, Los Angeles, CA 90048 USA. the other authors are with the Biomedical Imaging Research Institute · All authors are with Cedars-Sinai Medical Center, Los Angeles, CA 90048 USA. Xiuzhen Huang is with the Department of Computational Biomedicine
Deep-learning unrolled networks are the leading approach to accelerated MRI reconstruction. State-of-the-art models such as PromptMR+ learn a distinct block for every cascade, so their size grows with the number of cascades. Sharing one block across iterations avoids adding parameters, but earlier shared-weight models typically used a tiny block and produced worse images than models with distinct weights. Therefore, it remained unclear whether weight sharing itself or the choice of block was the problem. Recent advances in looped transformers, which reuse the same block across layers, show that they can match or outperform their distinct-weight counterparts on several tasks. Motivated by this, we propose LoopMR, which reuses a single Restormer block at every iteration as its core. Besides the current image estimate, the block receives data-consistency inputs and latent images. On four datasets covering the brain (CC359), knee (fastMRI), heart (CMRxRecon2023) and carotid (in-house), LoopMR significantly outperforms E2E-VarNet, PromptMR and PromptMR+ in PSNR and SSIM at every acceleration factor (p < 0.01), with about 1/9 of PromptMR+'s parameters (10.3M versus 92.9M). It also keeps its lead at unseen acceleration factors and on an unseen dataset. With our transformer block, the shared model matches or outperforms its distinct-weight counterpart at 6, 12 and 24 iterations. At a fixed number of iterations, shared transformer blocks match or outperform their distinct-weight versions, while shared CNN blocks perform worse. These results suggest that looping one well-chosen block can outperform stacking distinct cascades in MRI reconstruction, with a fraction of the parameters.
Figures & tables
Fig. 1: (a) Prior unrolled networks learn a new block at every cascade; LoopMR reuses one block at every iteration, with 1/T of the denoiser parameters. (b) SSIM (mean of R=5,10 ) against total parameters on CC359 brain.
Fig. 2: One LoopMR iteration. The shared Restormer receives the current estimate, its undersampled version, the zero-filled image, their residual and the latents, and returns an image update and new latents. Data consistency pulls the k-space toward the measurements. The same block is used at all T iterations.
Brain (CC359)
Knee (fastMRI)
×5
×10
×4
×8
Method
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
E2E-VarNet [ 12 ]
35.01 (0.42)
93.64 (0.59)
31.91 (0.47)
90.09 (0.87)
39.03 (1.83)
93.28 (2.80)
35.06 (2.33)
89.25 (3.93)
PromptMR [ 13 ]
36.96 (0.41)
94.94 (0.54)
34.29 (0.38)
92.98 (0.58)
40.13 (2.04)
93.95 (2.91)
36.84 (2.45)
90.94 (4.10)
PromptMR+ [ 14 ]
37.28 (0.41)
95.13 (0.54)
34.83 (0.35)
93.46 (0.55)
40.34 (2.08)
94.10 (2.92)
37.39 (2.48)
91.38 (4.14)
LoopMR (ours) ∗
37.45 (0.42)
95.28 (0.52)
35.12 (0.36)
93.77 (0.53)
40.44 (2.09)
94.24 (2.85)
37.54 (2.47)
91.61 (4.10)
TABLE I: Quantitative results on the brain (CC359) and knee (fastMRI) datasets.
Heart (CMRxRecon2023)
Carotid (in-house)
×4
×8
×10
×6
×12
Method
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
E2E-VarNet [ 12 ]
38.54 (2.80)
96.47 (1.87)
33.69 (2.63)
92.45 (3.14)
32.35 (2.51)
91.20 (3.33)
38.52 (2.53)
92.26 (3.41)
35.73 (2.58)
88.45 (4.44)
PromptMR [ 13 ]
41.34 (2.85)
97.76 (1.25)
36.27 (2.65)
94.93 (2.18)
34.83 (2.61)
94.00 (2.46)
39.39 (2.94)
93.01 (3.37)
37.43 (3.15)
90.32 (4.55)
PromptMR+ [ 14 ]
41.66 (2.77)
97.88 (1.17)
36.82 (2.63)
95.24 (2.12)
35.55 (2.65)
94.40 (2.49)
39.32 (2.91)
93.08 (3.34)
37.28 (3.09)
90.42 (4.49)
LoopMR (ours) ∗
41.93 (2.73)
98.00 (1.11)
37.25 (2.50)
95.59 (1.91)
36.00 (2.57)
94.85 (2.25)
40.16 (3.34)
93.48 (3.39)
38.04 (3.46)
90.91 (4.62)
TABLE II: Quantitative results on the heart (CMRxRecon2023) and carotid (in-house) datasets.
Method
Params (M)
Latency (ms)
Peak memory (GB)
E2E-VarNet
20.1–29.9
29–44
0.3–0.8
PromptMR
79.6–81.5
107–158
2.5–9.6
PromptMR+
92.9
100–121
1.0–3.3
LoopMR (ours)
10.3
203–624
0.7–3.0
TABLE III: Model size and inference cost over all four datasets, evaluated on one NVIDIA H100.
Fig. 3: Qualitative comparison on (a) brain (CC359, × 10), (b) knee (fastMRI, × 8) and (c) heart (CMRxRecon2023, × 10). Rows: reference, E2E-VarNet, PromptMR, PromptMR+ and LoopMR. For each dataset, the reconstruction (dashed box = ROI, enlarged in the inset) and its absolute error map are shown. Error maps: ∣x^−x∣ in each dataset’s own intensity units (colorbar below each dataset). Numbers below each image: per-slice SSIM / PSNR under each dataset’s reporting protocol.
Fig. 4: Qualitative comparison on the in-house carotid dataset (3-D MPRAGE, × 12, axial slice). Layout as in Fig. 3 , with error maps in the volume-max-normalized units of the MPRAGE evaluation.
Fig. 5: SSIM of the low-frequency (top row) and high-frequency (bottom row) components of the reconstructions by E2E-VarNet, PromptMR, PromptMR+ and LoopMR at the highest acceleration on (a) CC359 brain ( × 10), (b) fastMRI knee ( × 8), (c) CMRxRecon2023 heart ( × 10) and (d) in-house carotid ( × 12). As in [ 9 ] , the low-frequency component keeps the central third of k-space along each axis and the high-frequency component the rest. Boxes summarize the per-slice SSIM (median and quartiles), and whiskers extend to 1.5 × IQR. Outliers are not shown.
Fig. 6: Sharing weights across iterations, CC359 test set (10 volumes). (a) SSIM (mean of × 5 and × 10) against the number of iterations T for the shared model (10.2M in total, one 8.4M block) and the distinct models (52M, 102M, 202M in total), as means over volumes. (b) Distinct minus shared SSIM at × 5 and × 10 (mean and paired 95 % CI over volumes). Numbers: the number of volumes (out of 10) on which the shared model is better. (c) Mean distinct minus shared SSIM at × 10 for four blocks (two CNNs and two transformers), from a ≈ 2M block (open) to a 7.5–10M block (filled). Every pair in (c) is trained identically on both sides (12 iterations, full back-propagation, final-output loss, gradient clip 0.01) and evaluated at 12 iterations, and the 8.35M Restormer point is the matched T=12 pair of (a) and (b).
Block
Iterations
Params (M)
SSIM ×5
PSNR ×5
SSIM ×10
PSNR ×10
S better
Block family
(M)
S / D
S / D
S
D
S
D
S
D
S
D
( ×10 )
Depth
Restormer
8.35
6 / 6
10.2 / 52.0
0.9479
0.9477
36.69
36.68
0.9254
0.9236
33.92
33.77
10/10
Restormer
8.35
12 / 12
10.2 / 102.1
0.9484
0.9483
36.76
36.75
0.9265
0.9257
34.03
33.98
10/10
Restormer
8.35
24 / 24
10.2 / 202.3
0.9481
0.9481
36.74
36.73
0.9265
0.9251
34.05
33.94
10/10
Family
E2E-VarNet
2.46
12 / 12
2.9 / 30.0
0.9399
0.9434
35.51
36.00
0.9073
0.9142
32.38
32.98
0/10
E2E-VarNet
9.82
12 / 12
10.3 / 118.3
0.9431
0.9443
35.93
36.14
0.9141
0.9169
32.88
33.21
0/10
TABLE IV: Shared versus distinct weights on brain (CC359). S: one shared block. D: distinct blocks.
× 5
× 10
Params (M)
PSNR
SSIM
PSNR
SSIM
Image only
10.19
36.18
94.49
33.09
91.75
+ DC inputs
10.20
36.71
94.80
33.97
92.56
+ Latents
10.21
36.76
94.84
34.03
92.65
+ 5 adjacent slices
10.27
37.45
95.28
35.12
93.77
+ Latents, per-iterate loss
10.21
36.68
94.79
33.92
92.54
TABLE V: Ablation study of the proposed inputs and the loss of LoopMR on brain (CC359).
Fig. 7: Behavior of the loop over iterations, CC359 test set, × 10 (mean over the 10 test volumes). (a) SSIM and (b) PSNR of the estimate after iteration t . Final-output loss: the + Latents model of Table V . Per-iterate loss: the same model with a loss on every iteration. Both are trained with T=12 (dotted line) and run beyond it. PromptMR+: a single-slice PromptMR+ trained with the same schedule, after each of its 12 cascades (its depth is fixed at 12).
Unseen R (CC359)
Unseen data (fastMRI brain)
× 4
× 16
× 5
× 10
Method
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
E2E-VarNet [ 12 ]
35.31
93.96
29.37
85.91
31.33
84.27
29.00
79.15
PromptMR [ 13 ]
37.41
95.22
31.23
89.54
36.99
92.45
33.73
89.18
PromptMR+ [ 14 ]
37.73
95.40
32.10
90.45
37.30
92.48
34.57
89.33
LoopMR (ours)
37.91
95.58
33.02
91.75
38.15
93.09
35.40
90.14
TABLE VI: Ablation study of generalizability on unseen R and an unseen dataset.
Accelerated MRI reconstruction recovers diagnostic images from undersampled k-space data to reduce scan time while preserving anatomical and quantitative information.
Accelerated MRI reconstruction recovers images from undersampled k-space. However, different acceleration factors produce distinct artifact patterns. Existing methods often train separate models for each factor, leading to poor cross-factor generalization and high training and storage costs. We propose Shared LoRA, a parameter-efficient framework that freezes the pretrained SHFormer backbone and trains a single shared set of LoRA adapters together with a lightweight gating network. During training, undersampled inputs are generated by randomly sampling acceleration factors and their corresponding sampling masks, enabling the shared adapters to learn reconstruction knowledge across factors. Given the acceleration factor, GateNet generates layer-wise coefficients to dynamically modulate the residual strength of each adapter. Experiments show that Shared LoRA achieves the best or competitive PSNR and SSIM across acceleration factors, while its trainable parameters account for only about 5.3% of the total model parameters. Its performance at lower acceleration factors remains largely unaffected as the jointly trained factor set expands, and it generalizes stably to unseen neighboring factors.
Zhiwei Zhao, Weikang Gong, Zhongnian Li +1
School of Computer Science and Technology / School of Artificial Intelligence, China University of Mining and Technology, Xuzhou, China
Unrolled variational networks achieve state-of-the-art performance for accelerated MRI reconstruction, but standard end-to-end training becomes prohibitively memory intensive as the number and width of cascades increase. We propose a memory-efficient training framework that updates a randomly sampled cascade while retaining supervision from the final reconstruction loss. The error signal is propagated from the end of the chain to the sampled cascade using an activation-light backward pass that avoids storing full chain activations, thereby decoupling training memory from chain length. A time-conditioned network shared across cascades further keeps the parameter count independent of chain length. The backward pass is approximate only through frozen normalization statistics, and a trained exact gradient control shows the approximation leaves reconstruction quality unchanged. The framework remains trainable at chain lengths and cascade widths where standard end-to-end backpropagation exhausts GPU memory. On fastMRI single-coil knee and multi-coil brain reconstruction, our method consistently outperforms other local training approaches. On high-SNR knee slices, our method surpasses E2E-VarNet at 8× and matches it at 4×, while on the complete brain validation set it surpasses E2E-VarNet at 8× and achieves comparable performance at 4×, with fewer parameters, lower peak memory, and faster per-iteration training. Code is available at https://github.com/emehillaj/scalable-unrolled-mri.
Emiljo Mëhillaj, Ender Konukoglu
Computer Vision Lab, ETH Zurich, Zurich, Switzerland · LOOP Zurich - Medical Research Center, Zurich, Switzerland
Accelerated magnetic resonance imaging (MRI) forms images from incomplete Fourier measurements, shifting acquisition burden to computational reconstruction. Deep unrolled networks interleave learned regularization and data consistency, yet typically estimate the full unmeasured spectrum at every cascade. We propose STAIRS (Staging the Inverse Problem via Admissible-Support Scheduling), which projects intermediate reconstructions onto nested Fourier supports that expand to the full spectrum and conditions each update on its support transition. Conditional increments of a regularized log-determinant score of training covariance define a normalized spectral coordinate, accounting for dependencies between newly and previously admitted frequencies. Dyadic targets allocate larger coordinate increments to earlier cascades and are mapped to fixed supports before training, enabling coarse-to-fine reconstruction with unchanged sampling masks and data consistency. On fastMRI and Stanford2D, STAIRS improves matched Transformer (HUMUSNet) and convolutional (NAFNet) backbones in single- and multi-coil reconstruction, including radiologist-annotated pathology regions. It also outperforms predictor-based progression with lower parameter overhead and inference latency. Stage-wise analysis shows repeated error correction by projection and lower intermediate reconstruction errors than the backbones alone throughout support expansion. Ablations favor dyadic allocation over uniform expansion for square and line-band supports and over manual front-loading for square windows. Support projection alone improves reconstruction, and transition conditioning strengthens the gain. Cross-dataset evaluation shows improvements for both backbones in both directions. Support allocation thus complements regularizer design in unrolled MRI reconstruction.
Xiangjian Hou, Chang Ni, Chao Qin +6
Department of Radiology and Imaging Sciences and the Department of Electrical and Computer Engineering, University of Utah, Salt Lake City, UT 84112, USA · Department of Radiology and Imaging Sciences and the Department of Biomedical Engineering, University of Utah, Salt Lake City, UT 84112, USA · Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates +4