In this paper
AbstractIntroductionRelated WorkLearned Priors and Unrolled ReconstructionFourier Structure and Measurement DesignProgressive ReconstructionMethodAdmissible-Support UnrollingConditional Spectral CoordinateSchedule ConstructionOffline Computation and Host IntegrationExperimental ProtocolData and AcquisitionHosts and OptimizationResultsRealized Support SchedulesEffect of Support SchedulingMulti-coil ReconstructionCross-Dataset TransferSingle-Coil ReconstructionAblation StudiesDiscussionAppendixSensitivity of the Estimated Schedule to ReferencesWorksComputer Science
STAIRS: Staging the Inverse Problem via Admissible-Support Scheduling for Unrolled MRI Reconstruction
Abstract
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.
This work was supported in part by VA Merit Grant under I01CX003005.
(Corresponding author: Xiangjian Hou.)
Xiangjian Hou is with the Department of Radiology and Imaging Sciences and the Department of Electrical and Computer Engineering, University of Utah, Salt Lake City, UT 84112, USA (e-mail: [email protected]).
Chang Ni is with the Department of Radiology and Imaging Sciences and the Department of Biomedical Engineering, University of Utah, Salt Lake City, UT 84112, USA (e-mail: [email protected]).
Chao Qin is with the Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates (e-mail: [email protected]).
Zhiwei Tan and Xin Wang are with the Department of Electrical and Computer Engineering, University of Washington, Seattle, WA 98195, USA (e-mail: [email protected]; [email protected]).
Xinrui Ye and Rong-Rong Chen are with the Department of Electrical and Computer Engineering, University of Utah, Salt Lake City, UT 84112, USA (e-mail: [email protected]; [email protected]).
Chun Yuan and Xiaodong Ma are with the Department of Radiology and Imaging Sciences, University of Utah, Salt Lake City, UT 84112, USA (e-mail: [email protected]; [email protected]). Chun Yuan is also with the Department of Radiology, University of Washington, Seattle, WA 98195, USA (e-mail: [email protected]).
Introduction
Magnetic resonance imaging (MRI) offers rich soft-tissue contrast without ionizing radiation, but the spatial encoding required to form an image is inherently time consuming. Accelerated MRI shortens the examination by sampling only part of -space. Parallel imaging exploits the spatial diversity of receiver coils [1, 2], whereas compressed sensing exploits sparsity in a suitable transform domain to recover images from undersampled -space below the Nyquist rate [3]. Both approaches shift part of the acquisition burden toward computation, formulating image reconstruction as an inverse problem constrained by known encoding physics and undersampled data.
Deep unrolled reconstruction has become a natural computational model for this setting. An unrolled network maps a finite number of iterations of a model-based solver to a trainable architecture, interleaving learned regularization with explicit data consistency (DC) [4, 5, 6]. This combination of learned image priors and encoding physics has shown strong reconstruction performance on benchmarks such as fastMRI [7, 8], making the cascade a central unit of modern MRI reconstruction.
Most subsequent advances have refined the learned reconstruction modules or the information exchanged between cascades. Recurrent hidden states [9], recurrent multiscale Transformers [10], hybrid multiscale processing [11], shifted-window attention [12], prompting [13], and structured state-space modules [14] have increased the capacity of learned reconstruction models. Joint image and coil sensitivity estimation [15, 16] and cross-cascade feature fusion [17] have further enriched the reconstruction state. These developments largely retain the full Fourier support of the learned estimates: an early cascade can populate the same unmeasured coefficients as the final cascade. The intermediate reconstruction evolves with depth, while the learned estimates remain spectrally unrestricted.
The acquisition model fixes which measurements are available, but does not prescribe an identical support for every learned estimate. MRI provides a natural setting in which to examine this choice. Dependencies within local -space neighborhoods [18] and across coil channels [19, 20] enable structured low-rank reconstruction, while calibration methods exploit related structure to estimate coil sensitivities [21]. The nonuniform distribution of spectral energy also motivates variable-density sampling [3]. These properties motivate a statistical description of support expansion that accounts for both spectral variation and dependence, rather than the number or location of coefficients alone.
Progressive Divide-and-Conquer (PDAC) [22] uses a predictor to select stage-wise spectral supports, while flow-aligned training [23] shapes intermediate trajectories through cascade parameters and supervision. For support-based progression, the central scheduling question is when each spectral region enters the learned estimates. Early admission retains that region over more cascades for refinement, while later admission extends the preceding period of support-based filtering. Distributing these transitions across a finite cascade budget therefore determines the refinement stages available to different spectral regions. Equal increments in window width need not add equal amounts of spectral variation, and newly admitted coefficients can be strongly correlated with those already present. A schedule informed by the conditional contribution of each expansion connects this statistical structure to the allocation of reconstruction stages.
We propose STAIRS (Staging the Inverse Problem via Admissible-Support Scheduling), which constructs this allocation from the spectral statistics of fully sampled training targets. Each cascade applies a prescribed Fourier-support projection and conditions its learned update on the support transition. The supports are nested and expand to the full grid at the final cascade. Every DC step uses all acquired measurements, while the physical sampling mask and DC formulation remain unchanged. The complete schedule is computed before network training and remains fixed during training and inference, without an input-dependent support predictor.
Schedule construction uses a regularized log-determinant functional of training-set spectral covariance. On a prescribed nested family, its Schur-complement increments measure the contribution of newly admitted coefficients after accounting for their correlations with the existing support. Their normalized cumulative sum defines a monotone spectral coordinate along the family. Stage targets allocate this coordinate across cascades, and each target selects the smallest candidate support that attains it. We use dyadic targets, whose successive increments halve: early cascades admit a larger share of the coordinate, and later cascades add progressively smaller shares while continuing to refine the regions already admitted. We evaluate this allocation against uniform, back-loaded, and manually designed alternatives. Because the same targets are mapped through the coordinate of each MRI setting, one allocation rule yields a setting-specific support path, which both host regularizers share.
We instantiate STAIRS with Transformer-based HUMUS-Net [11] and an unrolled CNN with a NAFNet-based regularizer [24], and evaluate single- and multi-coil knee reconstruction on fastMRI [7] and Stanford2D [25] with multi-coil knee MRI. Matched-host comparisons show reconstruction gains with both hosts. For cross-dataset evaluation, we transfer the trained reconstructor and its schedule together, without fine-tuning or re-estimation. STAIRS improves over the matched baselines and, under the shared HUMUSNet host, over PDAC in both transfer directions. Under the shared HUMUSNet implementation, STAIRS also uses fewer parameters and has lower measured latency than PDAC.
Our contributions are as follows:
We introduce STAIRS, a framework for unrolled MRI reconstruction that couples progressive support projection along a fixed, data-derived schedule with transition conditioning, improving the intermediate reconstructions passed between cascades.
We characterize support expansion through conditional log-determinant increments and combine the resulting spectral coordinate with dyadic stage allocation, achieving higher reconstruction accuracy than uniform and manually designed schedules under a fixed backbone.
We show that STAIRS improves both convolutional and Transformer backbones, retaining its reconstruction gains when the source model and schedule are transferred unchanged across datasets. With HUMUSNet, it also outperforms PDAC while using fewer parameters and achieving lower inference latency.
Related Work
Learned Priors and Unrolled Reconstruction
End-to-end unrolling embeds learned regularization and explicit data consistency in a jointly trained reconstruction pipeline. ADMM-Net [26], variational networks [5], and MoDL [6] parameterize iterations of regularized inverse problems. Cascaded CNNs couple learned image corrections to data consistency [4], while KIKI-Net extends this processing across image and -space domains [27]. E2E-VarNet demonstrates the benefit of jointly learning sensitivity estimation and multi-coil reconstruction [8].
Within this formulation, recurrent state [9] carries information across iterations. Multiscale representations [11] and Transformer-based processing [10], [12] enrich the learned prior, while prompting [13] and structured state-space modules [14] provide contextual conditioning and long-range modeling, respectively. NAFNet shows that effective restoration priors can also be built from simple gated convolutions [24]. Joint image–sensitivity estimation [15], [16] and cross-cascade feature propagation [17] further extend the variables updated or exchanged between cascades. STAIRS retains this end-to-end formulation: the reconstruction modules are trained jointly along a fixed schedule, with one support projection per cascade. Schedule construction is independent of the regularizer architecture and complements the convolutional and Transformer hosts considered here.
Fourier Structure and Measurement Design
Structured spectral models exploit linear dependencies among Fourier coefficients to recover missing data. SAKE expresses these dependencies through low-rank block-Hankel matrices of multi-coil -space [19]. LORAKS relates limited image support and smooth phase to local low-rank structure [18], while P-LORAKS couples this structure across coils [20]. ALOHA connects transform-domain sparsity to annihilating-filter relations [28], and GIRAF exploits convolutional structure to solve related low-rank recovery problems efficiently [29]. ESPIRiT uses calibration subspaces to estimate coil sensitivities [21]. These methods use within-scan structure to constrain the reconstructed signal or estimate its encoding model. STAIRS shares the premise of spectral dependence, but uses covariance across training spectra to allocate stage supports, rather than imposing a structured low-rank constraint on each reconstruction.
Statistical criteria also guide acquisition design. Bayesian experimental design selects sampling patterns under a probabilistic reconstruction model [30], while OEDIPUS uses estimation bounds under sparsity constraints [31]. Learning-based methods optimize masks for a fixed reconstruction rule [32] or jointly optimize sampling and reconstruction [33]; sequential approaches adapt acquisition to observed data [34]. These approaches select the measurements available to a reconstructor. STAIRS instead uses spectral statistics to allocate projection supports after the acquisition mask has been fixed.
Progressive Reconstruction
PDAC [22] decomposes subsampling into nested reconstruction stages, combining hand-prescribed line-count budgets with input-dependent line selection and severity conditioning; it evaluates several hand-crafted budget schedules via enumeration, adopting the one that performs best empirically. STAIRS shares the use of nested supports and stage conditioning, but chooses stage supports by thresholding a correlation-aware training-set log-determinant coordinate within a prescribed family. The resulting sizes and locations are fixed before network training and shared across inputs and hosts.
UEPS [35] also expands a central Fourier region, but couples this expansion to increasing reconstruction resolution and incorporates acquired measurements as the grid grows. STAIRS keeps the reconstruction grid fixed and uses all acquired measurements in every DC step; the schedule changes the admissible support of each cascade, not the grid resolution.
FLAT [23] organizes cascade trajectories through flow-derived parameters and intermediate velocity supervision. STAIRS instead constrains Fourier support by projection at each cascade and retains the host reconstruction objective, without prescribing intermediate image or velocity targets.
Coarse-to-fine reconstruction also includes learned iterates on increasing-resolution grids [36], and pyramid recurrent reconstruction [37]. Related ideas include frequency continuation in waveform inversion [38], [39], frequency marching in cryo-EM [40], and the low-frequency bias of neural networks [41]. STAIRS uses a fixed reconstruction grid and computes the support allocation from a correlation-aware coordinate of training spectra.
Method
Admissible-Support Unrolling
Let denote the target image on a Cartesian Fourier grid . Multi-coil measurements follow
where is the receiver-sensitivity operator, is the Fourier transform, and restricts the data to the acquired index set . The single-coil model is recovered by setting . A conventional model-based reconstruction estimates from
with encoding prior information. Unrolled networks replace the optimizer of (2) by a finite sequence of learned updates and data-consistency operations derived from .
STAIRS makes the Fourier support of the learned estimate depend on cascade depth. We prescribe nested admissible supports
which specify the unmeasured coefficients admitted at each stage. At cascade , the acquired coefficients and the coefficients admitted by form the effective support . Its image-domain projection is
Writing for the support transition, the staged cascade is
The learned update and data-consistency map retain the form of the host network. The projection restricts the learned estimate to the stage-admissible support after DC incorporates all acquired measurements. Since , the terminal projection is the identity. Figure 1 summarizes the reconstruction cascade and support schedule.
Figure 1. STAIRS overview. Each cascade applies a learned update, physical data consistency, and an admissible-support projection on a fixed reconstruction grid. The support schedule is computed once from training spectra using a correlation-aware log-determinant score and remains fixed during training and inference. All acquired measurements are used at every DC step, and the final cascade admits the full Fourier grid.
Conditional Spectral Coordinate
We select the stage supports as from a prescribed family of nested candidates,
Here is the ACS window. Our default candidates are centered square windows of increasing even width , from to the full grid. The same construction also applies to nested bands of phase-encoding lines, which we evaluate in the schedule ablation.
All candidates use the same preprocessed training spectra from fully sampled targets. Define
We score each candidate using a common scale :
For independent and , this equals the mutual information : the information that a noisy observation of the coefficients in carries about a Gaussian surrogate of the training spectrum. Along the nested family, the increase in this score measures the contribution of each new shell after accounting for its dependence on the preceding support.
Proposition 1 (Conditional spectral increment). For candidates evaluated from the same centered samples and scale , each shell contributes
where
and denotes the empirical cross-covariance.
Proof. Write and , and group the coordinates of accordingly. Since , the block-determinant identity gives
Taking half the logarithm of both sides gives the equality in (10). Moreover, is the Schur complement of the -block in the positive semidefinite matrix . Thus , implying . Under the Gaussian model, the Gaussian conditioning formula identifies this same matrix with .
The conditional covariance removes the variation in explained by the noisy observation of the preceding support. A shell with large total variance can therefore contribute a small increment when its coefficients are strongly predictable from those already included. An energy criterion such as instead counts all variation in the shell, including this shared component. Normalizing the accumulated increments gives the spectral coordinate
since . By Proposition 1, , and is the fraction of the score beyond the reference support that remains outside . The remaining spectral score satisfies .
Schedule Construction
With the candidate family parameterized by , stage allocation reduces to selecting cumulative targets and mapping them back to supports. For a target , we use the first candidate that reaches it:
This rule reaches the requested spectral coordinate while admitting only the smallest support available in the candidate family.
Since is non-decreasing and the candidates are nested, admits the fewest unmeasured coefficients among the candidates with . By (12), its residual score satisfies
In the single-coil model, these unmeasured coefficients are exactly the coordinates on which vanishes, so the rule also admits the fewest unobserved directions. The inversion maps each target to a support, while the target sequence determines when successive spectral regions enter the learned estimate and how many later cascades are available to refine them.
We adopt a front-loaded allocation through normalized dyadic targets,
and set for , with . The target increments , where , halve at each cascade. This assigns larger shares of the spectral coordinate to earlier cascades, where admitted coefficients have more subsequent stages for refinement. Later targets add progressively smaller shares as the support approaches the full grid. Mapping these targets through determines the support sizes: front-loading is specified in the spectral coordinate, while the resulting widths depend on the training spectra. Discrete candidates can produce repeated indices, and assigning ensures full terminal support even when the score reaches a plateau earlier.
Offline Computation and Host Integration
To evaluate the scores without forming a covariance matrix in the full feature dimension, we use the sample-space Gram matrix
where contains as its rows. Sharing the nonzero eigenvalues of , this matrix gives the same score through . For the retained positive eigenvalues of , with , we set
Schedule sensitivity to this choice is examined in the Appendix.
Algorithm 1 Offline STAIRS schedule construction
Require: Training spectra ; nested candidates ; acquired set ; cascade count .
Ensure: Stage supports and effective supports .
Center the training features and form .
Compute by (15) and choose by (16).
for do
Form using (15).
.
end for
, .
for do
.
.
.
end for
.
, .
Algorithm 1 summarizes the offline schedule construction. The schedule is computed once from the training partition of each dataset and acquisition setting, shared across both hosts, and fixed during network training and inference. For cross-dataset evaluation, we retain the schedule from the source dataset.
In full STAIRS, transition conditioning makes the support change explicit to the learned update. The descriptor contains the preceding and current supports, identifying the region carried forward and the coefficients newly admitted. A mask encoder embeds for feature-wise affine modulation [42], applied to Transformer features in HUMUSNet [11] and normalized convolutional features in the NAFNet-based unrolled host [24]. The cascades are trained jointly using the host reconstruction objective along the precomputed schedule. The support-only control removes conditioning while retaining the projections; the conditioning-only control uses the same schedule descriptors with at every cascade.
Experimental Protocol
Data and Acquisition
We evaluate on the fastMRI knee benchmark [7] and Stanford2D FSE [25]. We use the 973 training and 199 validation volumes of fastMRI. Single-coil images are center-cropped to and reconstructed at and acceleration, using central fractions of and , respectively. Multi-coil images are cropped to and evaluated at acceleration with a central fraction.
Stanford2D contains 89 fully sampled volumes. We follow the fixed volume-level 80/20 split used by HUMUSNet [11] and PDAC [22], crop images to , and use the same acquisition setting. For each dataset and acceleration factor, the support schedule is computed exclusively from the corresponding training partition.
FastMRI+ [43] adds slice-level bounding boxes, drawn by a radiologist around pathology, to the fastMRI knee volumes. We use them to test whether the gains reach clinically relevant regions.
Hosts and Optimization
STAIRS is integrated with two unrolled hosts whose learned regularizers have different architectures: Transformer-based HUMUSNet [11] and convolutional NAFNet [24]. NAFNet supplies the convolutional learned update within the unrolled MRI network. For multi-coil data, both hosts adopt the sensitivity estimation and -space data consistency of E2E-VarNet [8]. Each uses eight cascades, an image-domain objective, and AdamW. HUMUSNet is trained for 40 epochs in the single-coil experiments and 50 epochs in the multi-coil experiments; its learning rate starts at and is reduced to over the final five epochs. NAFNet is trained for 20 and 50 epochs, respectively, with cosine decay from to . Batch sizes are four for fastMRI and one for Stanford2D. All runs use an NVIDIA H200 GPU.
Within each host, the baseline and STAIRS share the cascade count, data split, sampling masks, initialization, loss, optimizer, training schedule, and evaluation code. STAIRS adds only the scheduled support projections and transition conditioning described in Sec. III-D.
Results
Realized Support Schedules
Figure 2 shows how a shared set of dyadic targets is realized as distinct support paths. On the multi-coil grids, the first target () selects widths of 78 for Stanford2D and 90 for fastMRI. At the second cascade, the widths are 186 and 170, respectively. The reversal in ordering shows how the training-set spectral profiles shape support expansion across stages.
Figure 2. Empirical spectral coordinates and realized support schedules for four MRI settings. Curves show the spectral coordinate as a function of support width; markers indicate the selected stage supports. Insets magnify the region near full support.
| Method | fastMRI knee | Stanford2D | ||||
| PSNR↑ | SSIM↑ | NMSE↓ | PSNR↑ | SSIM↑ | NMSE↓ | |
| Zero-filled | 27.42 | 0.7000 | 0.0704 | 29.24 | 0.7175 | 0.0948 |
| U-Net [44] | 34.18 | 0.8601 | 0.0151 | 33.77 | 0.8953 | 0.0333 |
| E2E-VarNet [8] | 36.81 | 0.8873 | 0.0090 | 36.48 | 0.9220 | 0.0172 |
| HUMUSNet [11] | 36.84 | 0.8879 | 0.0092 | 37.36 | 0.9266 | 0.0183 |
| + PDAC [22] | 37.12 | 0.8905 | 0.0085 | 37.77 | 0.9269 | 0.0169 |
| + STAIRS | 37.28 | 0.8925 | 0.0082 | 37.96 | 0.9316 | 0.0163 |
| NAFNet [24] | 37.13 | 0.8911 | 0.0085 | 37.10 | 0.9261 | 0.0198 |
| + STAIRS | 37.33 | 0.8930 | 0.0081 | 37.22 | 0.9272 | 0.0190 |
Table 1. Quantitative results of accelerated multi-coil MRI reconstruction on fastMRI knee and Stanford2D. Best results within each host are in teal.
Across all four settings, the later supports cluster near the full grid as the increments in the spectral coordinate become smaller. The resulting schedules combine large early advances in the coordinate with finer support expansion in the later cascades, reaching full support at C8.
Effect of Support Scheduling
STAIRS maintains more accurate intermediate reconstructions than the baseline networks throughout support expansion (Figure 3(a)–(Figure 3c)). We evaluate the states passed to the next cascade using NMSE over all unmeasured complex coil -space coefficients. Across all three settings, STAIRS achieves lower mean intermediate NMSE than the baseline at every constrained stage C1–C7. With HUMUSNet, it also outperforms PDAC throughout C1–C7 on fastMRI and at six of the seven stages on Stanford2D.
Figure 3. Stage-wise reconstruction dynamics in multi-coil MRI. (a)–(c) Intermediate -space NMSE and (d)–(f) relative NMSE reduction after projection; error bars show pointwise 95% volume-bootstrap confidence intervals. (g), (h) Projection effects in effusion and ACL ROIs, with image-space NMSE reported below each reconstructed crop.
The projection gains reveal repeated error removal along these trajectories (Figure 3(d)–(Figure 3f)). Each gain is the relative NMSE reduction between the same prediction before and after projection. On fastMRI, STAIRS yields positive corrections throughout C1–C7 with both hosts. The corrections are strongest in the middle cascades, where the corresponding baselines develop broad error peaks. With HUMUSNet, STAIRS removes approximately 40%–70% of the prediction error across C2–C5, and its projection gain exceeds PDAC’s at every constrained stage. PDAC provides much smaller corrections and increases error at C1.
As the schedule admits more of the unmeasured spectrum in later cascades, the projection gains decrease while the intermediate states remain more accurate than those of the baseline networks. On Stanford2D, STAIRS exceeds PDAC’s projection gain throughout C4–C7, while its mean intermediate NMSE is approximately 20%–23% lower than PDAC’s at C5–C7 (Figure 3(b) and (Figure 3e)). At C8, full support makes projection the identity and its gain is zero.
These corrections are also visible in the fastMRI ROI examples. In the C3 effusion ROI with HUMUSNet, the STAIRS projection suppresses stripe-like artifacts and brings the curved fluid contour closer to the reference (Figure 3(g)). Image-space NMSE within this ROI falls by 84.6% after projection, compared with 0.45% for PDAC. With NAFNet, the C4 projection reduces residual structured artifacts around the ACL and lowers ROI NMSE by 37.9% (Figure 3(h)).
Multi-coil Reconstruction
STAIRS improves both hosts on both datasets in all three metrics, with PSNR gains of 0.12–0.60 dB (Table I). A host with STAIRS gives the best value in every column, and, under the shared HUMUSNet host, STAIRS also outperforms PDAC throughout. The gains with both Transformer and convolutional regularizers show that support allocation complements these two forms of learned regularization.
The gains extend to radiologist-annotated pathology (Table II). Pooled over all ROIs, STAIRS improves both hosts in all three metrics, and NAFNet+STAIRS is the best model; under HUMUSNet, the improvement holds in every anatomical group. PDAC behaves differently: its 0.28 dB image-level gain over HUMUSNet shrinks to 0.05 dB within the ROIs, and its SSIM falls below that of the host in every group, whereas STAIRS retains a 0.25 dB gain. Figure 4 shows representative reconstructions within fastMRI+ pathology ROIs, where STAIRS recovers the annotated structures with lower error than the baseline networks.

Figure 4. Qualitative comparison of 8× multi-coil reconstructions on fastMRI knee. Radiologist-annotated ROIs from fastMRI+ are magnified alongside their error maps. The top ROI shows the curved border of a joint effusion; the bottom ROI shows bone contours in a bone-injury case. Cyan arrows highlight local error reductions with STAIRS.
On Stanford2D (Figure 5), adding STAIRS to HUMUSNet improves the delineation of small bright structures in the upper example and reduces residual errors along the curved tissue boundaries in the lower example.

Figure 5. Qualitative comparison of multi-coil reconstructions on Stanford2D. The top ROI shows small high-signal structures and tissue borders; the bottom ROI shows curved tissue interfaces. Cyan arrows highlight local error reductions with STAIRS.
For multi-coil reconstruction with HUMUSNet, STAIRS uses 119.08M parameters, 15.9% fewer than PDAC’s 141.60M. With fixed inputs resident on the GPU, its model-only forward pass takes 185.79 ms, compared with 194.64 ms for PDAC and 177.94 ms for the baseline. At inference, STAIRS uses the precomputed support schedule directly, eliminating the auxiliary support-prediction network.
Cross-Dataset Transfer
For cross-dataset evaluation between fastMRI and Stanford2D, we transfer the trained reconstructor and its precomputed schedule together. STAIRS improves both hosts in every metric and direction, with PSNR gains of 0.15–1.06 dB, and HUMUSNet+STAIRS achieves the best results throughout (Table III). Under the shared HUMUSNet host, STAIRS exceeds PDAC by 1.36 dB for fastMRI-to-Stanford2D transfer and 1.52 dB in the reverse direction. PDAC improves HUMUSNet in-domain but falls below the baseline in both transfer directions. STAIRS thus retains its advantage over each host when the support schedule computed from source training spectra is used on the target dataset.
| Method | All ROIs () PSNR | All ROIs () SSIM | All ROIs () NMSE | Meniscus & cartilage () PSNR | Meniscus & cartilage () SSIM | Meniscus & cartilage () NMSE | Ligament & support () PSNR | Ligament & support () SSIM | Ligament & support () NMSE | Bone () PSNR | Bone () SSIM | Bone () NMSE | Joint & soft tissue () PSNR | Joint & soft tissue () SSIM | Joint & soft tissue () NMSE |
| HUMUSNet | 35.39 | 0.8922 | 0.0184 | 35.89 | 0.9131 | 0.0206 | 35.04 | 0.8739 | 0.0207 | 34.36 | 0.8196 | 0.0205 | 34.84 | 0.9003 | 0.0064 |
| + PDAC | 35.44 | 0.8911 | 0.0183 | 35.92 | 0.9121 | 0.0205 | 35.11 | 0.8733 | 0.0202 | 34.40 | 0.8162 | 0.0204 | 34.90 | 0.9001 | 0.0064 |
| + STAIRS | 35.64 | 0.8950 | 0.0174 | 36.12 | 0.9154 | 0.0196 | 35.33 | 0.8785 | 0.0191 | 34.58 | 0.8218 | 0.0194 | 35.11 | 0.9034 | 0.0061 |
| NAFNet | 35.51 | 0.8943 | 0.0181 | 35.98 | 0.9143 | 0.0203 | 35.16 | 0.8787 | 0.0194 | 34.52 | 0.8239 | 0.0198 | 35.02 | 0.9029 | 0.0062 |
| + STAIRS | 35.67 | 0.8957 | 0.0173 | 36.15 | 0.9160 | 0.0194 | 35.45 | 0.8800 | 0.0186 | 34.56 | 0.8222 | 0.0196 | 35.18 | 0.9043 | 0.0060 |
Table 2. ROI-BASED RESULTS OF MULTI-COIL RECONSTRUCTION WITHIN FASTMRI+ PATHOLOGY BOUNDING BOXES, POOLED AND PER ANATOMICAL GROUP. BEST RESULTS WITHIN EACH HOST ARE IN TEAL.
| Method | fastMRI → Stanford2D PSNR↑ | fastMRI → Stanford2D SSIM↑ | fastMRI → Stanford2D NMSE↓ | Stanford2D → fastMRI PSNR↑ | Stanford2D → fastMRI SSIM↑ | Stanford2D → fastMRI NMSE↓ |
| HUMUSNet | 34.14 | 0.8703 | 0.0382 | 33.62 | 0.8516 | 0.0171 |
| + PDAC | 33.57 | 0.8674 | 0.0443 | 32.25 | 0.8227 | 0.0232 |
| + STAIRS | 34.93 | 0.8901 | 0.0343 | 33.77 | 0.8557 | 0.0167 |
| NAFNet | 33.77 | 0.8830 | 0.0440 | 32.16 | 0.8287 | 0.0236 |
| + STAIRS | 34.50 | 0.8899 | 0.0368 | 33.22 | 0.8468 | 0.0188 |
Table 3. Cross-dataset evaluation of 8× multi-coil reconstruction (A→B: trained on A, tested on B). Best results within each host are in teal.
Single-Coil Reconstruction
Single-coil reconstruction sets in (1), which removes the coil sensitivities and their estimation. DC then modifies only the measured -space coefficients, so every unmeasured coefficient is recovered by the learned updates alone, from the cascade whose support first admits it. We therefore use this setting to test the schedule on the unrolled inverse problem itself. STAIRS improves both hosts at 4× and 8×, and NAFNet+STAIRS achieves the best results in every column (Table IV). Representative 8× reconstructions are shown in Figure 6. The gain of the schedule thus does not rely on coil sensitivities.

Figure 6. Qualitative comparison of single-coil reconstructions on fastMRI knee. The top ROI shows the lobulated outline of a soft-tissue lesion; the bottom ROI shows narrow tissue interfaces in a joint-effusion case. Cyan arrows highlight local error reductions with STAIRS.
Ablation Studies
Scheduled support and transition conditioning play complementary roles (Table V). Support projection alone raises PSNR from 37.36 to 37.55 dB and improves SSIM and
| Method | 4× PSNR↑ | 4× SSIM↑ | 4× NMSE↓ | 8× PSNR↑ | 8× SSIM↑ | 8× NMSE↓ |
| Zero-filled | 29.49 | 0.6541 | 0.0532 | 26.84 | 0.5500 | 0.0893 |
| U-Net [44] | 31.88 | 0.7142 | 0.0357 | 29.78 | 0.6424 | 0.0511 |
| D5C5 [4] | 32.25 | 0.7256 | 0.0332 | 29.65 | 0.6457 | 0.0512 |
| OUCR [45] | 32.61 | 0.7354 | 0.0315 | 30.59 | 0.6634 | 0.0443 |
| ReconFormer [10] | 32.73 | 0.7383 | 0.0310 | 30.89 | 0.6697 | 0.0429 |
| FMT-Net [46] | 32.55 | 0.7323 | 0.0423 | 30.99 | 0.6656 | 0.0536 |
| FPS-Former [47] | 32.51 | 0.7337 | 0.0316 | 31.03 | 0.6692 | 0.0408 |
| HUMUSNet [11] | 32.37 | 0.7221 | 0.0323 | 31.04 | 0.6722 | 0.0410 |
| + PDAC [22] | 32.73 | 0.7376 | 0.0310 | 31.16 | 0.6739 | 0.0407 |
| + STAIRS | 32.76 | 0.7394 | 0.0308 | 31.24 | 0.6786 | 0.0401 |
| NAFNet [24] | 32.74 | 0.7385 | 0.0309 | 31.22 | 0.6767 | 0.0402 |
| + STAIRS | 32.82 | 0.7411 | 0.0306 | 31.32 | 0.6799 | 0.0396 |
Table 4. Quantitative results of 4× and 8× accelerated single-coil MRI reconstruction on fastMRI knee. Best results within each host are in teal.
NMSE, without adding parameters. Conditioning the updates
| Variant | Support | Conditioning | PSNR↑ | SSIM↑ | NMSE↓ |
| HUMUSNet | 37.36 | 0.9266 | 0.0183 | ||
| Conditioning only | ✓ | 37.26 | 0.9144 | 0.0189 | |
| Support only | ✓ | 37.55 | 0.9282 | 0.0176 | |
| STAIRS | ✓ | ✓ | 37.96 | 0.9316 | 0.0163 |
Table 5. Ablation of support projection and transition conditioning (Sec. III-D) for 8× multi-coil reconstruction on Stanford2D with the HUMUSNet host. Best results are in teal.
on these enforced support transitions brings a further 0.41 dB gain, reaching 37.96 dB with the best SSIM and NMSE. The conditioning-only control has the same parameter count as full STAIRS but retains full support at every cascade; it achieves 37.26 dB and underperforms the host on all three metrics. The schedule changes the spectral task presented to each stage, and conditioning helps the learned update respond to that change.
We next compare how support expansion is allocated across cascades (Figure 7, Table VI). Within the square-window family, all target-based variants use the same spectral mapping, with uniform width as a geometric reference. Uniform width, uniform targets, and back-loaded targets all underperform the baseline on every metric, whereas both manual front-loaded allocations improve PSNR and NMSE over it. The comparison therefore favors admitting a larger share of the spectral coordinate early. How expansion is distributed within a front-loaded schedule also matters: relative to the aggressive front-loaded profile, STAIRS uses narrower supports at C1–C2, the same support at C3, and wider supports at C4–C7, and gains 0.43 dB PSNR. Its dyadic targets thus restrict the support more strongly than this profile at the start and refine a broader spectral region at C4–C7, yielding the best results on all three metrics among the square-window schedules.
| Schedule | PSNR | SSIM | NMSE |
| Baseline (full support) | 37.36 | 0.9266 | 0.0183 |
| Square windows | |||
| Uniform width | 37.23 | 0.9226 | 0.0192 |
| Back-loaded targets | 36.41 | 0.9196 | 0.0227 |
| Uniform targets | 36.96 | 0.9229 | 0.0204 |
| Moderate front-loaded | 37.51 | 0.9275 | 0.0181 |
| Aggressive front-loaded | 37.53 | 0.9261 | 0.0182 |
| STAIRS (dyadic targets) | 37.96 | 0.9316 | 0.0163 |
| Line bands | |||
| Uniform width | 36.41 | 0.9137 | 0.0226 |
| Uniform targets | 36.06 | 0.9125 | 0.0249 |
| STAIRS (dyadic targets) | 37.72 | 0.9294 | 0.0172 |
Table VI. Ablation of stage allocation with square and line-band supports for multi-coil reconstruction on Stanford2D with the HUMUSNet host. Best results within each support family are highlighted in teal.
Figure 7. Expansion of square supports under alternative allocations on Stanford2D MC8. Solid curves use spectral targets; the dashed uniform-width curve is a geometric reference.
We also evaluate the allocation on nested bands of phase-encoding lines (Table VI). With line-band supports, STAIRS reaches 37.72 dB PSNR, exceeding uniform-width and uniform-target schedules by 1.31 and 1.66 dB, respectively, with corresponding improvements in SSIM and NMSE. Both support families thus favor dyadic allocation over uniform expansion, with square windows yielding the best overall reconstruction.
Discussion
STAIRS makes the allocation of Fourier support across cascades a design variable in unrolled MRI reconstruction. The learned regularizer determines how an estimate is updated, and the support schedule determines which reconstructed spectral components are retained after data consistency. The gains with HUMUSNet and the NAFNet-based host in both single- and multi-coil reconstruction show that support allocation complements different regularizer architectures (Tables I and IV). The reconstruction path can thus be designed alongside the update that operates on it.
Support scheduling acts on the errors propagated between cascades. For HUMUSNet on fastMRI MC8, the scheduled projection reduces the error of the same network prediction at every constrained stage, and STAIRS maintains lower intermediate errors than the baseline and PDAC [22] (Figure 3). The benefit is visible in both the local effect of projection and the quality of the resulting intermediate estimates. In the ablation study, scheduled support improves reconstruction on its own, while transition conditioning strengthens the gain when paired with the scheduled projections (Table V). The support transition gives conditioning a concrete change in the reconstruction task to encode.
The schedule allocates refinement time to the unmeasured spectrum. Early admission allows reconstructed spectral content to remain in the projected estimates over more cascades, providing additional stages of refinement. Delayed admission prolongs the filtering imposed by the support restriction.
The schedule ablations favor dyadic allocation over uniform schedules for both square and line-band supports, and over aggressive front-loading for square windows (Table VI). Within each support family, the target-based variants share the same spectral coordinate, so their differences reveal the importance of allocation across depth. The conditional spectral coordinate converts this common allocation rule into support widths that follow the training spectra of each setting (Figure 2).
The cross-dataset results give the fixed schedule a practical role beyond in-domain reconstruction. STAIRS improves both hosts in both directions between fastMRI and Stanford2D with the source reconstructor and schedule transferred together (Table III). Under the shared HUMUSNet host, it also exceeds PDAC. A schedule derived from training spectral statistics can therefore remain effective across datasets, while the learned updates respond to the measurements of each scan. Both hosts use this precomputed schedule without an auxiliary support-prediction network.
The pooled fastMRI+ gains show that the benefit extends to annotated pathology regions (Table II), connecting support scheduling to fidelity at both image and regional scales. A natural next step is to optimize stage targets jointly with cascade depth, using the conditional spectral coordinate to distribute refinement across the spectrum. Whole-image and ROI evaluations provide complementary criteria for selecting those allocations.
Appendix
Sensitivity of the Estimated Schedule to
The spectral coordinate yields consistent support schedules across alternative lower-tail estimates of (Table VII). Replacing the median in (16) with the mean leaves every stage support unchanged, while the 25th and 75th percentiles change the selected widths by at most four pixels.
| Median | Mean | 25th pct. | 75th pct. | |
| () | 4.825 | 4.668 | 3.707 | 5.695 |
| C1 | 78 | 78 | 76 | 78 |
| C2 | 186 | 186 | 182 | 190 |
| C3 | 268 | 268 | 266 | 270 |
| C4 | 318 | 318 | 316 | 320 |
| C5 | 362 | 362 | 360 | 362 |
| C6 | 374 | 374 | 374 | 374 |
| C7 | 382 | 382 | 380 | 382 |
| C8 | 384 | 384 | 384 | 384 |
| 0 | 0 | 4 | 4 |
Table VII. Sensitivity of the Stanford2D multi-coil schedule to the lower-tail estimate of . Stage widths and their maximum absolute changes from the default median schedule are in pixels.
References
- [1]K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boesiger, “SENSE: sensitivity encoding for fast MRI,” Magnetic Resonance in Medicine, vol. 42, no. 5, pp. 952–962, 1999.pubmed.ncbi.nlm.nih.gov/10542355
- [2]M. A. Griswold, P. M. Jakob, R. M. Heidemann, M. Nittka, V. Jellus, J. Wang, B. Kiefer, and A. Haase, “Generalized autocalibrating partially parallel acquisitions (GRAPPA),” Magnetic Resonance in Medicine, vol. 47, no. 6, pp. 1202–1210, 2002.pubmed.ncbi.nlm.nih.gov/12111967
- [3]M. Lustig, D. Donoho, and J. M. Pauly, “Sparse MRI: The application of compressed sensing for rapid MR imaging,” Magnetic Resonance in Medicine, vol. 58, no. 6, pp. 1122–1195, 2007.pubmed.ncbi.nlm.nih.gov/17969013
- [4]J. Schlemper, J. Caballero, J. V. Hajnal, A. N. Price, and D. Rueckert, “A deep cascade of convolutional neural networks for dynamic MR image reconstruction,” IEEE Transactions on Medical Imaging, vol. 37, no. 2, pp. 491–503, Feb 2018. [Online]. Available: http://dx.doi.org/10.1109/TMI.2017.2769078
- [5]K. Hammernik, T. Klatzer, E. Kobler, M. P. Recht, D. K. Sodickson, T. Pock, and F. Knoll, “Learning a variational network for reconstruction of accelerated MRI data,” Magnetic Resonance in Medicine, vol. 79, no. 6, pp. 3055–3071, 2018.arxiv.org/abs/1704.00447
- [6]H. K. Aggarwal, M. P. Mani, and M. Jacob, “MoDL: Model-based deep learning architecture for inverse problems,” IEEE Transactions on Medical Imaging, vol. 38, no. 2, pp. 394–405, Feb 2019. [Online]. Available: http://dx.doi.org/10.1109/TMI.2018.2865356
- [7]J. Zbontar, F. Knoll, A. S. Jin, T. Murrell, Z. Huang, M. J. Muckley, A. Defazio, R. Stern, P. Johnson, M. Bruno et al., “fastMRI: An open dataset and benchmarks for accelerated MRI,” arXiv preprint arXiv:1811.08839, 2018.arxiv.org/abs/1811.08839
- [8]A. Sriram, J. Zbontar, T. Murrell, A. Defazio, C. L. Zitnick, N. Yakubova, F. Knoll, and P. Johnson, “End-to-end variational networks for accelerated MR image reconstruction,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2020, pp. 64–73.DOI
- [9]G. Yiasemis, J.-J. Sonke, C. Sánchez, and J. Teuwen, “Recurrent variational network: A deep learning inverse problem solver applied to the task of accelerated MRI image reconstruction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 732–741.arxiv.org/abs/2111.09639
- [10]P. Guo, Y. Mei, J. Zhou, S. Jiang, and V. M. Patel, “ReconFormer: Accelerated MRI reconstruction using recurrent transformer,” IEEE Transactions on Medical Imaging, vol. 43, no. 1, pp. 582–593, Jan 2024. [Online]. Available: http://dx.doi.org/10.1109/TMI.2023.3314747
- [11]Z. Fabian, B. Tinaz, and M. Soltanolkotabi, “HUMUS-Net: Hybrid unrolled multi-scale network architecture for accelerated MRI reconstruction,” in Advances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Associates, Inc., 2022, pp. 25 306–25 319. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2022/file/a1bb3f96e2655ae1e04325ae166bce0f-Paper-Conference.pdfarxiv.org/abs/2203.08213
- [12]J. Huang, Y. Fang, Y. Wu, H. Wu, Z. Gao, Y. Li, J. Del Ser, J. Xia, and G. Yang, “Swin transformer for fast MRI,” Neurocomputing, vol. 493, pp. 281–304, 2022.arxiv.org/abs/2201.03230
- [13]B. Xin, M. Ye, L. Axel, and D. N. Metaxas, “Fill the k-space and refine the image: Prompting for dynamic and multi-contrast MRI reconstruction,” in Statistical Atlases and Computational Models of the Heart. Regular and CMRxRecon Challenge Papers, O. Camara, E. Puyol-Antón, M. Sermesant, A. Suniesiaputra, Q. Tao, C. Wang, and A. Young, Eds. Cham: Springer Nature Switzerland, 2024, pp. 261–273.arxiv.org/abs/2309.13839
- [14]Y. Korkmaz and V. M. Patel, “MambaRecon: MRI reconstruction with structured state space models,” in Proceedings of the Winter Conference on Applications of Computer Vision (WACV), February 2025, pp. 4142–4152.arxiv.org/abs/2409.12401
- [15]M. Arvinte, S. Vishwanath, A. H. Tewfik, and J. I. Tamir, “Deep J-Sense: Accelerated MRI reconstruction via unrolled alternating optimization,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2021, M. de Bruijne, P. C. Cattin, S. Cotin, N. Padoy, S. Speidel, Y. Zheng, and C. Essert, Eds. Cham: Springer International Publishing, 2021, pp. 350–360.arxiv.org/abs/2103.02087
- [16]Y. Jun, H. Shin, T. Eo, and D. Hwang, “Joint deep model-based MR image and coil sensitivity reconstruction network (Joint-ICNet) for fast MRI,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 5266–5275.DOI
- [17]B. Xin, M. Ye, L. Axel, and D. N. Metaxas, “Rethinking deep unrolled model for accelerated MRI reconstruction,” in Computer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds. Cham: Springer Nature Switzerland, 2025, pp. 164–181.DOI
- [18]J. P. Haldar, “Low-rank modeling of local k-space neighborhoods (LO-RAKS) for constrained MRI,” IEEE Transactions on Medical Imaging, vol. 33, no. 3, pp. 668–681, 2014.
- [19]J. P. Shin, P. E. Larson, M. A. Ohliger, M. Elad, J. M. Pauly, D. B. Vigneron, and M. Lustig, “Calibrationless parallel imaging reconstruction based on structured low-rank matrix completion,” Magnetic Resonance in Medicine, vol. 72, no. 4, pp. 959–970, 2014.DOI
- [20]J. P. Haldar and J. Zhuo, “P-LORAKS: low-rank modeling of local k-space neighborhoods with parallel imaging data,” Magnetic Resonance in Medicine, vol. 75, no. 4, pp. 1499–1514, 2016.DOI
- [21]M. Uecker, P. Lai, M. J. Murphy, P. Virtue, M. Elad, J. M. Pauly, S. S. Vasanawala, and M. Lustig, “ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA,” Magnetic Resonance in Medicine, vol. 71, no. 3, pp. 990–1001, 2014.DOI
- [22]C. Wang, L. Guo, Y. Wang, H. Cheng, Y. Yu, and B. Wen, “Progressive divide-and-conquer via subsampling decomposition for accelerated MRI,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 25 128–25 137.arxiv.org/abs/2403.10064
- [23]K. Qi, S. Gupta, X. Hu, Q. Hu, W. Lyu, Y. Wang, and C. Chen, “FLAT: FLow-Aligned Training of unrolled networks for MRI reconstruction,” 2026. [Online]. Available: https://arxiv.org/abs/2512.03020
- [24]L. Chen, X. Chu, X. Zhang, and J. Sun, “Simple baselines for image restoration,” in European Conference on Computer Vision. Springer, 2022, pp. 17–33.arxiv.org/abs/2204.04676
- [25]J. Y. Cheng, “Stanford 2D FSE,” http://mridata.org/list?project=Stanford+2D+FSE, accessed: 2026-02-10.
- [26]Y. Yang, J. Sun, H. Li, and Z. Xu, “Deep ADMM-Net for compressive sensing MRI,” in Advances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, Eds., vol. 29. Curran Associates, Inc., 2016. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2016/file/1679091c5a880faf6fb5e6087eb1b2dc-Paper.pdf
- [27]T. Eo, Y. Jun, T. Kim, J. Jang, H.-J. Lee, and D. Hwang, “KIKI-Net: cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images,” Magnetic Resonance in Medicine, vol. 80, no. 5, pp. 2188–2201, 2018.DOI
- [28]K. H. Jin, D. Lee, and J. C. Ye, “A general framework for compressed sensing and parallel MRI using annihilating filter based low-rank Hankel matrix,” IEEE Transactions on Computational Imaging, vol. 2, no. 4, pp. 480–495, 2016.arxiv.org/abs/1504.00532
- [29]G. Ongie and M. Jacob, “A fast algorithm for convolutional structured low-rank matrix recovery,” IEEE Transactions on Computational Imaging, vol. 3, no. 4, pp. 535–550, 2017.
- [30]M. Seeger, H. Nickisch, R. Pohmann, and B. Schölkopf, “Optimization of k-space trajectories for compressed sensing by Bayesian experimental design,” Magnetic Resonance in Medicine, vol. 63, no. 1, pp. 116–126, 2010.DOI
- [31]J. P. Haldar and D. Kim, “OEDIPUS: An experiment design framework for sparsity-constrained MRI,” IEEE Transactions on Medical Imaging, vol. 38, no. 7, pp. 1545–1558, 2019.
- [32]B. Gözcü, R. K. Mahabadi, Y.-H. Li, E. Ilıcak, T. Cukur, J. Scarlett, and V. Cevher, “Learning-based compressive MRI,” IEEE Transactions on Medical Imaging, vol. 37, no. 6, pp. 1394–1406, 2018.
- [33]C. D. Bahadiri, A. Q. Wang, A. V. Dalca, and M. R. Sabuncu, “Deep-learning-based optimization of the under-sampling pattern in MRI,” IEEE Transactions on Computational Imaging, vol. 6, pp. 1139–1152, 2020. [Online]. Available: http://dx.doi.org/10.1109/TCI.2020.300627
- [34]T. Yin, Z. Wu, H. Sun, A. V. Dalca, Y. Yue, and K. L. Bouman, “End-to-end sequential sampling and reconstruction for MRI,” in Proceedings of Machine Learning for Health, ser. Proceedings of Machine Learning Research, S. Roy, S. Pfohl, E. Rocheteau, G. A. Tadesse, L. Oala, F. Falck, Y. Zhou, L. Shen, G. Zamzmi, P. Mugambi, A. Zirikly, M. B. A. McDermott, and E. Alsentzer, Eds., vol. 158, PMLR, 04 Dec 2021, pp. 261–281. [Online]. Available: https://proceedings.mlr.press/v158/yin21a.htmlarxiv.org/abs/2105.06460
- [35]X. Zhou, H. Shang, Z. Zhan, T. He, J. Meng, and D. Liang, “UEPS: Robust and efficient MRI reconstruction,” 2026. [Online]. Available: https://arxiv.org/abs/2603.18572
- [36]A. Hauptmann, J. Adler, S. Arridge, and O. Öktem, “Multi-scale learned iterative reconstruction,” IEEE Transactions on Computational Imaging, vol. 6, pp. 843–856, 2020.DOI
- [37]E. Z. Chen, P. Wang, X. Chen, T. Chen, and S. Sun, “Pyramid convolutional RNN for MRI image reconstruction,” IEEE Transactions on Medical Imaging, vol. 41, no. 8, pp. 2033–2047, 2022.
- [38]C. Bunks, F. M. Saleck, S. Zaleski, and G. Chavent, “Multiscale seismic waveform inversion,” Geophysics, vol. 60, no. 5, pp. 1457–1473, 1995.DOI
- [39]L. Sirgue and R. G. Pratt, “Efficient waveform inversion and imaging: A strategy for selecting temporal frequencies,” Geophysics, vol. 69, no. 1, pp. 231–248, 2004.
- [40]A. Barnett, L. Greengard, A. Pataki, and M. Spivak, “Rapid solution of the cryo-EM reconstruction problem by frequency marching,” SIAM Journal on Imaging Sciences, vol. 10, no. 3, pp. 1170–1195, 2017.arxiv.org/abs/1610.00404
- [41]N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. A. Hamprecht, Y. Bengio, and A. Courville, “On the spectral bias of neural networks,” in Proceedings of the 36th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 97. PMLR, 2019, pp. 5301–5310.arxiv.org/abs/1806.08734
- [42]E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville, “FiLM: Visual reasoning with a general conditioning layer,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, Apr 2018. [Online]. Available: http://dx.doi.org/10.1609/aaai.v32i1.11671
- [43]R. Zhao, B. Yaman, Y. Zhang, R. Stewart, A. Dixon, F. Knoll, Z. Huang, Y. W. Lui, M. S. Hansen, and M. P. Lungren, “fastMRI+, clinical pathology annotations for knee and brain fully sampled magnetic resonance imaging data,” Scientific Data, vol. 9, no. 1, p. 152, 2022.DOI
- [44]O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2015, pp. 234–241.arxiv.org/abs/1505.04597
- [45]P. Guo, J. M. J. Valanarasu, P. Wang, J. Zhou, S. N. Jiang, and V. M. Patel, “Over-and-under complete convolutional RNN for MRI reconstruction,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2021, pp. 13–23.arxiv.org/abs/2106.08886
- [46]Q. Yi, F. Fang, G. Zhang, and T. Zeng, “Frequency learning via multi-scale Fourier transformer for MRI reconstruction,” IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 11, pp. 5506–5517, 2023.DOI
- [47]Y. Meng, Z. Yang, Y. Shi, and Z. Song, “Boosting ViT-based MRI reconstruction from the perspectives of frequency modulation, spatial purification, and scale diversification,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 6, 2025, pp. 6135–6143.



