Pith. sign in

REVIEW 4 major objections 5 minor 65 references

SHFormer: Dynamic Spectral Filtering Convolutional Neural Network and High-pass Kernel Generation Transformer for Adaptive MRI Reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read The paper proposes SHFormer, a hybrid MRI reconstruction network that uses spectral filtering in a CNN and dynamic high-pass kernel generation in a transformer to recover high-frequency detail and generalize to unseen contrasts and accelera

desk verdict Solid architecture paper whose headline open-set generalization claim is weakened by a subject-overlap confound. read the letter →

arxiv 2607.20159 v1 pith:OYU6BE3C submitted 2026-07-22 cs.CV

classification cs.CV
keywords MRIreconstructionattentionmechanismspectralfilteringhigh-passkernelsdomaingeneralizationself-supervisedlearningdiffusionmodelsneuromodulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that attention-based MRI reconstruction can be made both sharper and more portable by performing attention in the frequency domain rather than only in the image domain. It introduces SHFormer, whose CNN branch recalibrates features through learned DCT spectral attention and whose transformer branch generates per-image high-pass kernels to emphasize edges and fine structures. The authors argue this addresses two known weaknesses of attention in MRI: bias toward low frequencies, which produces smooth reconstructions, and the need to retrain per contrast or anatomy. They report consistent gains across supervised, self-supervised, and diffusion-model training, with the largest margins in closed- and open-set generalization experiments. If correct, a single model could serve multiple MRI contrasts and acceleration factors without retraining.

What carries the argument

Two cooperating attention modules. The spectral filtering CNN applies the discrete cosine transform along channel and spatial dimensions, keeps top-K coefficients, and learns attention masks that reweight frequency components per input, intended to capture mode-specific transferable features. The dynamic high-pass kernel generation transformer uses a small hypernetwork to predict per-location low-pass kernels, converts them to high-pass kernels, and applies them to self-attention features, forcing high-frequency detail through the network.

What would settle it

Train SHFormer on several contrasts from a single scanner and test on an unseen contrast from the same scanner; if the ~1 dB PSNR and ~0.01 SSIM gains over baselines shrink or vanish, the open-set generalization is at least partly explained by dataset shift. Equivalently, hold the contrast fixed and vary the scanner to see whether the method transfers across acquisition sites.

Watch

Extended reading notes

Core claim

The paper claims that a neuromodulation-based, multi-spectral attention mechanism—implemented as a DCT-based spectral filtering CNN plus a dynamic high-pass kernel generation transformer—lets one MRI reconstruction model both propagate high-frequency detail and carry features reusable across deviated unseen domains. In their experiments, this design outperforms CNNs, GANs, transformers, adaptive weight-prediction networks, and meta-learning baselines on cardiac, knee, and multi-contrast brain data, including settings where the test contrast was unseen during training. The authors also show the modules can be dropped into a measurement-conditioned diffusion model and a self-supervised physics

Load-bearing premise

The open-set generalization claim assumes that the only meaningful distribution shift in the leave-one-out experiments is the MRI contrast, not differences in scanner, acquisition protocol, or preprocessing between source and target datasets; if those confounds dominate, the experiments do not isolate modality transfer.

Editorial extensions

If this is right

  • A single trained model can reconstruct cardiac and knee MRI across multiple acceleration factors and mask patterns with consistent gains over dedicated baselines.
  • The same architectural modules improve a diffusion-based reconstruction model with only about 1% more parameters and 0.14% more FLOPs.
  • In closed-set generalization, the network handles unseen acceleration factors with a flatter performance drop at high factors, suggesting reduced need for retraining.
  • In open-set generalization to unseen contrasts, the model gains roughly 1 dB PSNR and 0.01 SSIM over adaptive baselines, and fine-tuning only the small hypernetworks recovers most of the adaptation benefit.
  • Ablation studies indicate both modules matter: spectral filtering adds about 0.3 dB PSNR, while high-pass kernel generation adds over 0.01 SSIM.
  • The design transfers to self-supervised learning, where no fully sampled ground truth is available.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the frequency-domain mechanism is what drives transfer, similar spectral attention modules could be dropped into other inverse problems with varying acquisition physics, such as PET or CT reconstruction.
  • The open-set gains may partly reflect dataset or scanner differences rather than pure contrast transfer; isolating scanner effects would sharpen the claim.
  • The finding that fine-tuning only 2-3% of parameters (the two hypernetworks) restores most adaptation performance suggests a practical deployment recipe for new sites or contrasts.
  • A natural next test is whether the model generalizes to contrasts not in the DCT top-K set, which would probe whether the spectral attention truly learns modality-invariant components.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SHFormer, a cascaded MRI reconstruction architecture combining a DCT-based spectral filtering CNN (SFCNN) and a dynamic high-pass kernel generation transformer (HKTFB). The authors motivate the design by the low-frequency bias of attention mechanisms and the need for models that adapt to heterogeneous MRI contrasts and undersampling patterns. They evaluate SHFormer in supervised single-coil reconstruction (ACDC cardiac, fastMRI knee), physics-driven self-supervised multi-coil reconstruction, diffusion-model-based reconstruction (MC-DDPM with SH modules), closed-set generalization across unseen acceleration factors, open-set generalization across unseen MRI contrasts, and ablations. The paper reports consistent improvements over a wide range of baselines, with headline gains of roughly 1 dB PSNR and 0.01 SSIM under unseen scenarios, and releases code at GitHub.

Significance. If the results are validated, SHFormer would be a competitive and broadly applicable reconstruction architecture. Its combination of DCT-based spectral filtering with dynamic high-pass kernels is a plausible mechanism for improving high-frequency recovery and cross-contrast adaptability, and the breadth of experiments—supervised, self-supervised, diffusion, closed/open-set generalization, ablations, and visualizations—is a strength. The availability of code supports reproducibility. However, the central 'unseen modality' claim is currently undermined by a subject-overlap confound in the open-set protocol, and the closed-set protocol contains inconsistencies about which acceleration factors are truly unseen. These issues must be resolved before the paper's main contribution can be assessed.

major comments (4)
  1. [Sec. 4.1, Table 6] The leave-one-out open-set protocol does not control for subject identity. MRBrainS supplies T1 and FLAIR from the same 7 volumes, and IXI supplies T2 and PD from the same 14 volumes. Thus in 'T1, FL, PD→T2', the model trains on PD from the exact IXI volumes whose T2 is the test target; in 'T1, T2, PD→FL' it trains on T1 from the MRBrainS volumes whose FLAIR is tested. The reported ~1 dB open-set gains can therefore be explained by subject-specific anatomical priors rather than modality-invariant feature reuse. A subject-disjoint split, or explicit evidence that no subject appears in both train and test contrasts, is required. The same concern applies to Table 7, where the 'Unseen' PD→PDFS experiment does not state whether PDFS test volumes overlap with PD training volumes.
  2. [Sec. 4.3.4, Table 5] The protocol description and the table conflict on what is 'unseen'. The text states training on AFs 2x, 2.5x, 3.3x, 4x, 5x, and 8x, then evaluating on 'unseen' AFs from 2x to 10x. Table 5, however, reports seven AFs as 'unseen' while 5x is in the training list, and the text also mentions '35 unseen acceleration factors' even though the grid from 2x to 10x includes seen values. This ambiguity makes it impossible to determine which closed-set results are at truly held-out accelerations and can inflate apparent generalization. Please re-state the exact train/test AF split and exclude seen AFs from 'unseen' reporting.
  3. [Sec. 4.3.2, Table 2] The diffusion-model comparison contains an internal numerical inconsistency. The text reports a best PSNR improvement of 0.14 dB for MRBrains 5x, but Table 2 gives 34.68 → 34.86, a 0.18 dB gain. Also, the 'Computation cost analysis' refers to 'Table 4.3.1' rather than Table 2. Since the diffusion experiment is one of the three learning modes highlighted in the abstract, these numbers need to be reconciled.
  4. [Tables 1, 3, 4, 5, 6, 7] Many of the reported improvements over the strongest baselines are small (0.05–0.5 dB PSNR; 0.004–0.01 SSIM) while the reported standard deviations are often much larger. No paired significance tests, confidence intervals, or per-volume comparisons are provided. Given the paper's claim of 'consistent and superior performance', such tests would substantially strengthen the empirical case; without them, marginal differences (e.g., knee 8x: 30.85 vs 30.80 dB) may not be reliable. This is not fatal by itself, but it is load-bearing for the headline conclusion.
minor comments (5)
  1. [Table 1] The baseline 'OUCR []' has an empty citation; a reference is needed.
  2. [Tables 3 and 6] Table 3 contains the typo 'IST A' for ISTA-Net; Table 6 has an entry formatted as '38.50 /- 0.9713' which appears to be a typographical error.
  3. [Sec. 4.1] The dataset name is spelled inconsistently as 'MRBrains' and 'MRBrainS'; please unify.
  4. [Sec. 4.3.4, Figures 12/13] The text says '35 unseen acceleration factors', Figure captions say '15 unseen cases', and Table 5 lists 7 factors; please clarify what is plotted and what the 35 refers to.
  5. [Sec. 4.3.5] The term 'open-set generalization' is nonstandard for covariate shift between image contrasts; consider using 'cross-contrast generalization' or explicitly defining the term in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SHFormer's claims are empirical architecture results evaluated against external and prior baselines, not derivations from fitted constants or self-citations.

full rationale

SHFormer is an MRI reconstruction architecture, and its reported improvements are empirical benchmark results. The method is defined directly in Eqs. 7-19 and optimized with L1 loss; no reported PSNR/SSIM value is a fitted parameter renamed as a prediction, and no quantity used for evaluation is obtained from the model's own output by construction. The closed-set generalization tests unseen acceleration factors and mask patterns, and the open-set tests leave out entire contrasts, so those results are not forced by training on the same targets. Self-citations such as MAC-ReconNet, MCI-HyperNet, KM-MAML, and SFT-KD-Recon appear as baselines or related prior work, not as load-bearing justification for the proposed method's performance, and no uniqueness theorem from the authors' prior work is invoked to declare the architecture forced. The ablation selecting 16 DCT components is hyperparameter selection, not a circular derivation. The subject-overlap concern raised for Tables 6-7 is an experimental validity issue about whether open-set generalization is isolated from subject identity, but it does not show that any prediction reduces to its input by definition, so it does not raise the circularity score under the stated criteria.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The central claim is empirical; the main postulated content is the channel-DCT spectral-axis assumption and the adopted dynamic-HPF construction. Everything else is standard inverse-problem or diffusion background, or trainable parameters.

free parameters (4)
  • Top-K DCT components (K1/K2) = 16
    The spectral attention module truncates DCT coefficients; the paper selects 16 after a PSNR/SSIM sweep and uses it for all experiments. The transferability claim depends on this truncation not discarding informative frequencies.
  • Number of cascades N_c = 5
    Stated in the implementation details; depth is a manual design choice affecting reconstruction quality and capacity.
  • HPF kernel size K and group count G = not fully specified (K approximate 3 per Figure 6; G unspecified)
    The dynamic high-pass module's capacity depends on these values, but exact numbers are not reported in the main text.
  • Diffusion variance schedule and timesteps = not stated in main text
    The diffusion-based results rely on the schedule and number of timesteps, presumably in the supplementary; the main text does not provide enough detail.
assumptions (7)
  • standard math DCT transforms in Eqs. 7, 9, and 11 are orthonormal/invertible with the stated scaling.
    The spectral filtering module relies on taking top-K coefficients, applying attention, and inverting the DCT; invertibility is required for feature maps to remain meaningful.
  • domain assumption MRI acquisition is modeled by the under-sampled Fourier forward model in Eq. 1.
    Standard inverse-problem formulation for accelerated MRI; assumed throughout the paper.
  • domain assumption Self-supervised partition of k-space into y1/y2 with L1 loss (Eqs. 4-5) gives a valid training signal.
    Adopted from SSDU [53]; the self-supervised experiments depend on this.
  • domain assumption MC-DDPM formulation in Eq. 6 is valid for measurement-conditioned diffusion.
    Adopted from [51]; the diffusion variant rides on this prior result.
  • domain assumption DNN training first fits low-frequency components (F-principle [52]).
    Used to explain why the spectral filtering block enhances low frequencies and why high-frequency emphasis is needed.
  • ad hoc to paper Treating the channel dimension at each spatial location as a 1D signal for DCT (Eq. 7) captures meaningful spectral structure of feature maps.
    This is an unusual modeling choice; the paper does not justify why channels form a frequency axis. The spectral filtering module's transferability claim rests on it.
  • domain assumption LPF-to-HPF conversion by subtracting 1 from the middle kernel element (Eq. 18) creates a valid high-pass filter.
    Adopted from MMCA [25]; used to build the dynamic high-pass kernels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SHFormer: Dynamic Spectral Filtering Convolutional Neural Network and High-pass Kernel Generation Transformer for Adaptive MRI Reconstruction." pith.science (2026). https://pith.science/paper/OYU6BE3C

@misc{pith2026260720159,
  author       = {Pith},
  title        = {Pith review of: SHFormer: Dynamic Spectral Filtering Convolutional Neural Network and High-pass Kernel Generation Transformer for Adaptive MRI Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OYU6BE3C}},
  note         = {Machine review of arXiv:2607.20159}
}
read the original abstract

Attention Mechanism (AM) selectively focuses on essential information for imaging tasks and captures relationships between distant pixel neighborhoods to compute feature representations. Accelerated MRI reconstruction benefits from AM, as the imaging process involves Fourier domain measurements that influence image representation non-locally. However, AM-based models are more adept at capturing low-frequency information with limited capacity for high-frequency representations, restricting models to smooth reconstruction. Additionally, AM-based models need mode-specific retraining for multimodal MRI data, as their knowledge is restricted to local contextual variations that may be inadequate to capture transferable features across heterogeneous domains. To address these challenges, we propose a neuromodulation-based discriminative multi-spectral AM for scalable MRI reconstruction that can (i) propagate context-aware high-frequency details for high-quality reconstruction, and (ii) capture features reusable across deviated unseen domains in multimodal MRI. The proposed network consists of a spectral filtering CNN to capture mode-specific transferable features and a dynamic high-pass kernel generation transformer focusing on high-frequency details. We evaluate our model on comparative studies in supervised and self-supervised learning, diffusion model-based training, closed-set and open-set generalization under heterogeneous MRI data, and interpretation-based analysis. Our method offers scalable, high-quality reconstruction with best improvement margins of ~1 dB in PSNR and ~0.01 in SSIM under unseen scenarios. Code: https://github.com/sriprabhar/SHFormer

Figures

Figures reproduced from arXiv: 2607.20159 by the authors.

Figure 1
Figure 1. Concept diagram comparing previous neuromodulation-based AMs with the proposed method for heterogeneous MRI data. Methods like decoupled learning [33], coupled learning [32], and kernel modulation meta-learning [34] use a dynamic weight prediction (DWP) network with the backbone image reconstruction network. These methods extract task-specific features, wherein each mode of the multimodal MRI data is posed as a task… view at source ↗
Figure 2
Figure 2. AM in MRI reconstruction. (a) Input image (b) k-space of (a). The lower k-space frequencies correspond to smooth variations (red box) in the image while higher k-space frequencies correspond to fine details (cyan arrows). Distant patches with similar patterns share frequency components (cyan boxes in (a) and (b). (c) High-pass filtering AM features. (d) DCT-based spectral AM features. The AM computes the hidden repr… view at source ↗
Figure 3
Figure 3. A tabulated summary of various neuromodulation-based AMs (coupled learning [32], modulation MAML [34], adaptive instance normalization (Ada-IN) [24], channel attention (RCAN) [60], and deep frequency filtering (DFF) [22]) and the proposed AM approach for MRI reconstruction comparing various aspects - the scope of attention and adaptability, the type of interactions, and frequency perspectives Sriprabha Ramanarayanan… view at source ↗
Figures from the paper (19 more)
Figure 4
Figure 4. Figure 4: (Top Left) Architecture block diagram of SHFormer in the deep cascaded mode. (Bottom) Detailed network structure SHFormer. The spectral filtering module in the CNN extracts a spatial attention map that captures mode-specific reusable features of multimodal data. The dy…
Figure 5
Figure 5. Figure 5: Spectral filtering CNN (SFCNN) structure: The SFCNN consists of two DCT-based spectral AM stages operating on the encoder features of the CNN. The first stage learns a spatial frequency domain attention map by applying forward and inverse DCT along channel direction (𝑓…
Figure 6
Figure 6. Figure 6: Dynamic High-pass kernel generation (HK) block consists of a DWP network that predicts instance-specific low-pass filter kernels along each channel group for each spatial location of the features. The low-pass filter kernels are converted to high-pass filter kernels an…
Figure 7
Figure 7. Figure 7: Qualitative comparison of the predictions of SHFormer with other CNN-based, GAN-based, transformer-based, and KD-based methods for 5x cardiac MRI reconstruction. The yellow arrows show that the visual quality of the proposed model is much better for SHFormer than other…
Figure 8
Figure 8. Figure 8: Qualitative comparison of the predictions of SHFormer with other reconstructed images. The figures show the predictions for 8x knee MRI reconstruction. The quality of the proposed model is better than other reconstruction methods as highlighted near the ligament region…
Figure 9
Figure 9. Figure 9: Qualitative comparison of the predictions of MC-DDPM with and without the SF and HF layers for 5x under-sampling in MRBrains - axial brain T1 MRI (left), ACDC cardiac short axis view (middle), and Calgary brain image (right). In all three cases, the highlighted pattern…
Figure 10
Figure 10. Figure 10: Qualitative comparison of the predictions of SHFormer with those of other methods for physics-driven self-supervised 4x axial T2 knee multi-coil MRI reconstruction. The proposed method recovers textural details much closer to the target than the competitive baselines …
Figure 11
Figure 11. Figure 11: Qualitative comparison of the predictions of SHFormer with those of other methods for physics-driven self-supervised 4x coronal PD knee MRI multi-coil MRI reconstruction. The highlighted region shows the superior recovery of textural details on the bone regions for th…
Figure 12
Figure 12. Figure 12: Plot comparing the PSNR metrics for the closed-set generalization capabilities of SHFormer with that of joint training, other adaptive learning networks, and meta-learning approaches for several unseen acceleration factors from 4x to 10x when trained on a few accelera…
Figure 13
Figure 13. Figure 13: Plot comparing the SSIM metrics for the closed-set generalization capabilities of SHFormer with that of joint training, other adaptive learning networks, and meta-learning approaches for several unseen acceleration factors from 4x to 10x when trained on a few accelera…
Figure 14
Figure 14. Figure 14: Qualitative comparison of the closed-set generalization capabilities of SHFormer with joint training, task-level adaptive approaches at the architecture, and model-agnostic meta-learning methods for cardiac 8x undersampling. The results show higher accuracy for SHForm…
Figure 15
Figure 15. Figure 15: Qualitative comparison of the open-set generalization capabilities of SHFormer with other adaptive reconstruction models trained on T1, T2, and FLAIR and evaluated on PD. Here PD show much finer changes in the intensity level which are much better in the predictions o…
Figure 16
Figure 16. Figure 16: Qualitative comparison of the open-set generalization capabilities of SHFormer with other adaptive reconstruction models trained on T1, PD, and FLAIR and evaluated on T2. The residual images and the highlighted regions show improved recovery of faint structures for th…
Figure 17
Figure 17. Figure 17: Qualitative comparison of the open-set generalization capabilities of SHFormer with other adaptive reconstruction models for complex-valued fastMRI dataset with coronal PD and PDFS modalities. As pointed out by yellow arrows, the highlighted regions indicate better re…
Figure 18
Figure 18. Figure 18: Qualitative comparison of SHFormer with other multimodal integrated methods for T2 FLAIR MRI reconstruc￾tion. Sriprabha Ramanarayanan et al.: Preprint submitted to Elsevier Page 22 of 28 [PITH_FULL_IMAGE:figures/full_fig_p023_18.png]
Figure 19
Figure 19. Figure 19: Ablative study of the SHFormer showing the contributions of various functional units of the architecture towards the overall reconstruction performance. The visual results indicated by the highlighted regions clearly show the importance of the design that combines the…
Figure 20
Figure 20. Figure 20: Ablative study on varying numbers of significant DCT components in the spectral attention module. The performance increases as the number of frequency components increase and saturates beyond a certain extent, indicating redundancy in the parameters. Sriprabha Ramanar…
Figure 21
Figure 21. Figure 21: Visualization of the learned DCT-based attention maps for spectral attention. The DCT coefficients at the top left position are dense and correspond to low and mid frequency components while the maps at the lower right are sparse and indicate high frequency components…
Figure 22
Figure 22. Figure 22: (a) Visualization of the low pass filtered features (top) and features of the dynamic high pass filtering block (bottom). The features show higher response in regions where the intensity transitions are sharp, indicating that the dynamic kernels learns to filter high-…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 3 canonical work pages

  1. [1]

    Deep Learning Techniques for Automatic MRI CardiacMulti-StructuresSegmentationandDiagnosis:IstheProblem Solved? IEEE Trans

    Bernard,O.,Lalande,A.,Zotti,C.,Cervenansky,F.,Yang,X.,Heng, P., Cetin, I., Lekadir, K., Camara, O., Gonzalez Ballester, M.A., Sanroma,G.,Napel,S.,Petersen,S.,Tziritas,G.,Grinias,E.,Khened, M., Kollerathu, V.A., Krishnamurthi, G., Rohé, M., Pennec, X., Sermesant, M., Isensee, F., Jäger, P., Maier-Hein, K.H., Full, P.M., Wolf, I., Engelhardt, S., Baumgartne...

  2. [2]

    Controllable Image Processing via Adaptive FilterBank Pyramid

    Chen, D., Fan, Q., Liao, J., Aviles-Rivero, A., Yuan, L., Yu, N., Hua, G., 2020. Controllable Image Processing via Adaptive FilterBank Pyramid. IEEE Trans. Image Process. 29, 8043–8054. doi:10.1109/ TIP.2020.3009844

  3. [3]

    Neuralizer:Generalneuroimageanaly- sis without re-training, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Czolbe,S.,Dalca,A.V.,2023. Neuralizer:Generalneuroimageanaly- sis without re-training, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6217– 6230

  4. [4]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy,A.,Beyer,L.,Kolesnikov,A.,Weissenborn,D.,Zhai,X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N., 2020. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR abs/2010.11929. URL:https://arxiv.org/abs/2010.11929,arXiv:2010.11929

  5. [5]

    VS-Net: Variable Splitting Network for Accelerated Parallel MRI Reconstruction, in: Proc

    Duan,J.,Schlemper,J.,Qin,C.,Ouyang,C.,Bai,W.,Biffi,C.,Bello, G., Statton, B., O’Regan, D.P., Rueckert, D., 2019. VS-Net: Variable Splitting Network for Accelerated Parallel MRI Reconstruction, in: Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Intervent., pp. 713–722

  6. [6]

    Kiki- net: cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images

    Eo, T., Jun, Y., Kim, T., Jang, J., Lee, H.J., Hwang, D., 2018. Kiki- net: cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images. Magnetic resonance in medicine 80, 2188–2201

  7. [7]

    A General Decoupled Learning Framework for Parameterized Image Operators

    Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B., 2019. A General Decoupled Learning Framework for Parameterized Image Operators. IEEE Trans. Pattern Anal. Mach. Intell. , 1–1doi:10.1109/ TPAMI.2019.2925793

  8. [8]

    Multi-contrastmrisuper- resolution via a multi-stage integration network, in: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C

    Feng,C.M.,Fu,H.,Yuan,S.,Xu,Y.,2021a. Multi-contrastmrisuper- resolution via a multi-stage integration network, in: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (Eds.), Medical Image Computing and Computer Assisted Interven- tion – MICCAI 2021, Springer International Publishing, Cham. pp. 140–149

Show all 65 references
  1. [9]

    Feng, C.M., Yan, Y., Chen, G., Xu, Y., Hu, Y., Shao, L., Fu, H.,

  2. [10]

    (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2021, Springer International Publishing, Cham

    Feng,C.M.,Yan,Y.,Fu,H.,Chen,L.,Xu,Y.,2021b.Tasktransformer networkforjointmrireconstructionandsuper-resolution,in:deBrui- jne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI...

  3. [11]

    Model-Agnostic Meta- Learning for Fast Adaptation of Deep Networks, in: Proc

    Finn, C., Abbeel, P., Levine, S., 2017. Model-Agnostic Meta- Learning for Fast Adaptation of Deep Networks, in: Proc. ICML, pp. 1126–1135. URL:http://proceedings.mlr.press/v70/finn17a.html

  4. [12]

    Erasing concepts from diffusion models, in: Proceedings of the 2023 IEEE International Conference on Computer Vision

    Gandikota, R., Materzyńska, J., Fiotto-Kaufman, J., Bau, D., 2023. Erasing concepts from diffusion models, in: Proceedings of the 2023 IEEE International Conference on Computer Vision

  5. [13]

    G.S., R., Ramnarayanan, S., Fahim, M.A., Ram, K., S.P, P., Sivaprakasam, M., 2023. Sdlformer: A sparse and dense locality- enhanced transformer for accelerated mr image reconstruction, in: Xue, Z., Antani, S., Zamzmi, G., Yang, F., Rajaraman, S., Huang, S.X., Linguraru, M.G., ...

  6. [14]

    Hypernetworks

    Ha, D., Dai, A., Le, Q.V., 2016. Hypernetworks. arXiv preprint arXiv:1609.09106

  7. [15]

    Learning a Varia- tional Network for Reconstruction of Accelerated MRI Data

    Hammernik, K., Klatzer, T., Kobler, E., Recht, M.P., Sodick- son, D.K., Pock, T., Knoll, F., 2018. Learning a Varia- tional Network for Reconstruction of Accelerated MRI Data. Magn. Reson. Med. 79, 3055–3071. doi:10.1002/mrm.26977, arXiv:https://onlinelibrary.wiley.com/doi/pdf...

  8. [16]

    Denoising diffusion probabilistic models.arXiv:2006.11239

    Ho, J., Jain, A., Abbeel, P., 2020. Denoising diffusion probabilistic models.arXiv:2006.11239

  9. [17]

    Meta-learning in neural networks: A survey

    Hospedales, T.M., Antoniou, A., Micaelli, P., Storkey, A.J., 2021. Meta-learning in neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence , 1–1doi:10.1109/TPAMI. 2021.3079209

  10. [18]

    Hu, C., Li, C., Wang, H., Liu, Q., Zheng, H., Wang, S., 2021. Self- supervised learning for mri reconstruction with a parallel network training framework, in: Medical Image Computing and Computer AssistedIntervention–MICCAI2021:24thInternationalConference, Strasbourg,France,Se...

  11. [19]

    Swin transformer for fast mri

    Huang, J., Fang, Y., Wu, Y., Wu, H., Gao, Z., Li, Y., Del Ser, J., Xia, J., Yang, G., 2022. Swin transformer for fast mri. Neurocomputing 493, 281–304

  12. [20]

    MRI Reconstruction Via Cascaded Channel-Wise Attention Network, in: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), pp

    Huang, Q., Yang, D., Wu, P., Qu, H., Yi, J., Metaxas, D., 2019. MRI Reconstruction Via Cascaded Channel-Wise Attention Network, in: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), pp. 1622–1626. doi:10.1109/ISBI.2019.8759423

  13. [21]

    Li, A., Zhang, L., Liu, Y., Zhu, C., 2023. Feature modulation trans- former:Cross-refinementofglobalrepresentationviahigh-frequency prior for image super-resolution, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 12514– 12524

  14. [22]

    Deep frequency filteringfordomaingeneralization,in:ProceedingsoftheIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Lin,S.,Zhang,Z.,Huang,Z.,Lu,Y.,Lan,C.,Chu,P.,You,Q.,Wang, J., Liu, Z., Parulkar, A., Navkal, V., Chen, Z., 2023. Deep frequency filteringfordomaingeneralization,in:ProceedingsoftheIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11797–11807

  15. [23]

    Glid:Pre-trainingageneralist encoder-decoder vision model, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Liu,J.,Zheng,J.,Liu,Y.,Li,H.,2024. Glid:Pre-trainingageneralist encoder-decoder vision model, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 22851–22860

  16. [24]

    Universalundersampled mri reconstruction, in: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C

    Liu,X.,Wang,J.,Liu,F.,Zhou,S.K.,2021. Universalundersampled mri reconstruction, in: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (Eds.), Medical Image Comput- ing and Computer Assisted Intervention – MICCAI 2021, Springer International...

  17. [25]

    Dynamichigh-passfilteringandmulti-spectralattentionfor imagesuper-resolution,in:2021IEEE/CVFInternationalConference onComputerVision(ICCV),pp.4268–4277

    Magid, S.A., Zhang, Y., Wei, D., Jang, W.D., Lin, Z., Fu, Y., Pfister, H.,2021. Dynamichigh-passfilteringandmulti-spectralattentionfor imagesuper-resolution,in:2021IEEE/CVFInternationalConference onComputerVision(ICCV),pp.4268–4277. doi:10.1109/ICCV48922. 2021.00425

  18. [26]

    Cellular, synaptic and network effects of neuromodulation

    Marder, E., Thirumalai, V., 2002. Cellular, synaptic and network effects of neuromodulation. Neural Networks 15, 479–493. URL:https://www.sciencedirect.com/science/article/ pii/S0893608002000436, doi:https://doi.org/10.1016/S0893-6080(02) 00043-6

  19. [27]

    Matcha, N., Ramanarayanan, S., Fahim, M.A., S, R.G., Ram, K., Sivaprakasam, M., 2024. Sft-kd-recon: Learning a student-friendly teacherforknowledgedistillationinmagneticresonanceimagerecon- struction, in: Oguz, I., Noble, J., Li, X., Styner, M., Baumgartner, C., Rusu,M.,Heinma...

  20. [28]

    MRBrainS challenge: Online Evaluation Framework for Brain Image Segmentation in 3T MRI scans

    Mendrik, A.M., Vincken, K.L., Kuijf, H.J., Breeuwer, M., Bouvy, W.H.,DeBresser,J.,Alansary,A.,DeBruijne,M.,Carass,A.,El-Baz, A., et al., 2015. MRBrainS challenge: Online Evaluation Framework for Brain Image Segmentation in 3T MRI scans. Comput. Intell. Neurosci. 2015, 1

  21. [29]

    A review on the at- tention mechanism of deep learning

    Niu, Z., Zhong, G., Yu, H., 2021. A review on the at- tention mechanism of deep learning. Neurocomputing 452, 48–62. URL:https://www.sciencedirect.com/science/article/ pii/S092523122100477X, doi:https://doi.org/10.1016/j.neucom.2021. 03.091

  22. [30]

    Generalizabledeeplearningmethodforsuppressingunseenandmul- tiple mri artifacts using meta-learning.arXiv:2304.06378

    Palla, A., Ramanarayanan, S., Ram, K., Sivaprakasam, M., 2023. Generalizabledeeplearningmethodforsuppressingunseenandmul- tiple mri artifacts using meta-learning.arXiv:2304.06378

  23. [32]

    Mci-hypernet: A multi- ple contextual information-based adaptive weight learning net- work for controllable image reconstruction

    Ramanarayanan, S., Murugesan, B., Palla, A., Ram, K., Venkate- san, R., Sivaprakasam, M., 2023a. Mci-hypernet: A multi- ple contextual information-based adaptive weight learning net- work for controllable image reconstruction. Neurocomputing 554, 126606. URL:https://www.scienc...

  24. [33]

    Ramanarayanan, S., Murugesan, B., Ram, K., Sivaprakasam, M.,

  25. [34]

    Generalizingsuperviseddeeplearningmrireconstructiontomultiple and unseen contrasts using meta-learning hypernetworks

    Ramanarayanan, S., Palla, A., Ram, K., Sivaprakasam, M., 2023b. Generalizingsuperviseddeeplearningmrireconstructiontomultiple and unseen contrasts using meta-learning hypernetworks. Applied Soft Computing 146, 110633. URL:https://www.sciencedirect. com/science/article/pii/S156...

  26. [35]

    U-Net: Convolutional Networks for Biomedical Image Segmentation, in: Proc

    Ronneberger, O., Fischer, P., Brox, T., 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation, in: Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Intervent., pp. 234–241

  27. [36]

    Palette: Image-to-image diffusion models

    Saharia, C., Chan, W., Chang, H., Lee, C.A., Ho, J., Salimans, T., Fleet, D.J., Norouzi, M., 2021. Palette: Image-to-image diffusion models. CoRR abs/2111.05826. URL:https://arxiv.org/abs/2111. 05826,arXiv:2111.05826

  28. [37]

    Schlemper, J., Caballero, J., Hajnal, J.V., Price, A., Rueckert, D.,

  29. [38]

    Brain-inspired learning in artificial neural networks: a review.arXiv:2305.11252

    Schmidgall, S., Achterberg, J., Miconi, T., Kirsch, L., Ziaei, R., Hajiseyedrazi, S.P., Eshraghian, J., 2023. Brain-inspired learning in artificial neural networks: a review.arXiv:2305.11252

  30. [39]

    A hybrid, dual domain, cascade of convolutional neural networks for magnetic resonance image reconstruction, in: Proc

    Souza, R., Lebel, R.M., Frayne, R., 2019. A hybrid, dual domain, cascade of convolutional neural networks for magnetic resonance image reconstruction, in: Proc. Int. Conf. Med. Imaging Deep Learn

  31. [40]

    An open, multi-vendor, multi-field-strength brain mr dataset and analysis of publicly available skull stripping methods agreement

    Souza, R., Lucena, O., Garrafa, J., Gobbi, D., Saluzzi, M., Ap- penzeller, S., Rittner, L., Frayne, R., Lotufo, R., 2017. An open, multi-vendor, multi-field-strength brain mr dataset and analysis of publicly available skull stripping methods agreement. NeuroImage

  32. [41]

    Taskswitchingnetworkformulti-task learning, in: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Sun, G., Probst, T., Pani Paudel, D., Popovi, N., Kanakis, M., Patel, J.,Dai,D.,VanGool,L.,2021. Taskswitchingnetworkformulti-task learning, in: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 8271–8280. doi:10.1109/ICCV48922.2021.00818

  33. [42]

    JointCS-MRI Reconstruction and Segmentation with a Unified Deep Network, in: Proc

    Sun,L.,Fan,Z.,Ding,X.,Huang,Y.,Paisley,J.,2019. JointCS-MRI Reconstruction and Segmentation with a Unified Deep Network, in: Proc. Inf. Process. Med. Imaging, pp. 492–504

  34. [43]

    Compressed sensing mri using a recursive dilated network, in: Proc

    Sun, L., Fan, Z., Huang, Y., Ding, X., Paisley, J., 2018. Compressed sensing mri using a recursive dilated network, in: Proc. AAAI Conf. Artif. Intell

  35. [44]

    On-the-fly test-time adaptation for medical image segmentation

    Valanarasu, J.M.J., Guo, P., VS, V., Patel, V.M., 2022. On-the-fly test-time adaptation for medical image segmentation. URL:https: //arxiv.org/abs/2203.05574, doi:10.48550/ARXIV.2203.05574

  36. [45]

    Attention is all you need, in: Guyon, I., Luxburg, U.V., Bengio, S., Wal- lach, H., Fergus, R., Vishwanathan, S., Garnett, R

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I., 2017. Attention is all you need, in: Guyon, I., Luxburg, U.V., Bengio, S., Wal- lach, H., Fergus, R., Vishwanathan, S., Garnett, R. (Eds.), Ad- vancesinNeuralInformationP...

  37. [46]

    Multimodal model- agnostic meta-learning via task-aware modulation, in: Neural Infor- mation Processing Systems

    Vuorio, R., Sun, S.H., Hu, H., Lim, J.J., 2019. Multimodal model- agnostic meta-learning via task-aware modulation, in: Neural Infor- mation Processing Systems

  38. [47]

    Tent: Fully test-time adaptation by entropy minimization, in: Inter- national Conference on Learning Representations

    Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T., 2021. Tent: Fully test-time adaptation by entropy minimization, in: Inter- national Conference on Learning Representations. URL:https:// openreview.net/forum?id=uXl3bZLkr3c

  39. [48]

    Wang, H., Wu, X., Huang, Z., Xing, E.P., 2020. High-frequency component helps explain the generalization of convolutional neural networks, in: 2020 IEEE/CVF Conference on Computer Vision and PatternRecognition(CVPR),pp.8681–8691. doi:10.1109/CVPR42600. 2020.00871

  40. [49]

    CFSNet: Toward a Controllable Feature Space for Image Restoration, in: IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Wang, W., Guo, R., Tian, Y., Yang, W., 2019. CFSNet: Toward a Controllable Feature Space for Image Restoration, in: IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4139–

  41. [50]

    A deep ensemble network for compressed sensing mri, in: Proc

    Wu, H., Wu, Y., Sun, L., Cai, C., Huang, Y., Ding, X., 2018. A deep ensemble network for compressed sensing mri, in: Proc. Neural Inf. Process., pp. 162–171

  42. [51]

    Measurement-conditioned denoising diffusion probabilisticmodelforunder-sampledmedicalimagereconstruction

    Xie, Y., Li, Q., 2022. Measurement-conditioned denoising diffusion probabilisticmodelforunder-sampledmedicalimagereconstruction. arXiv:2203.03623

  43. [52]

    Training behavior of deep neural network in frequency domain, in: Gedeon, T., Wong, K.W., Lee,M.(Eds.),NeuralInformationProcessing,SpringerInternational Publishing, Cham

    Xu, Z.Q.J., Zhang, Y., Xiao, Y., 2019. Training behavior of deep neural network in frequency domain, in: Gedeon, T., Wong, K.W., Lee,M.(Eds.),NeuralInformationProcessing,SpringerInternational Publishing, Cham. pp. 264–274

  44. [53]

    Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data

    Yaman, B., Hosseini, S.A.H., Moeller, S., Ellermann, J., Uğurbil, K., Akçakaya, M., 2020. Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data. Magnetic resonance in medicine 84, 3172–3191

  45. [54]

    DAGAN:Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction

    Yang, G., Yu, S., Dong, H., Slabaugh, G., Dragotti, P.L., Ye, X., Liu, F.,Arridge,S.,Keegan,J.,Guo,Y.,Firmin,D.,2018. DAGAN:Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction. IEEE Trans. Med. Imaging 37, 1310–

  46. [55]

    The neural basis of selective attention

    Yantis, S., 2008. The neural basis of selective attention. Current Di- rections in Psychological Science 17, 86–90. URL:https://doi.org/ 10.1111%2Fj.1467-8721.2008.00554.x, doi:10.1111/j.1467-8721.2008. 00554.x

  47. [56]

    Yiasemis, G., Sonke, J.J., Sánchez, C., Teuwen, J., 2022. Recurrent variational network: a deep learning inverse problem solver applied to the task of accelerated mri reconstruction, in: Proceedings of the Sriprabha Ramanarayanan et al.:Preprint submitted to ElsevierPage 27 of...

  48. [57]

    Restormer: Efficient transformer for high-resolution image restoration, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., 2022. Restormer: Efficient transformer for high-resolution image restoration, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5728–5739

  49. [58]

    fastMRI: An open dataset and benchmarks for accelerated MRI

    Zbontar,J.,Knoll,F.,Sriram,A.,Muckley,M.J.,Bruno,M.,Defazio, A., Parente, M., Geras, K.J., Katsnelson, J., Chandarana, H., Zhang, Z., Drozdzal, M., Romero, A., Rabbat, M., Vincent, P., Pinkerton, J., Wang, D., Yakubova, N., Owens, E., Zitnick, C.L., Recht, M.P., Sodickson, D.K...

  50. [59]

    Ista-net: Interpretable optimization- inspireddeepnetworkforimagecompressivesensing,in:Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Zhang, J., Ghanem, B., 2018. Ista-net: Interpretable optimization- inspireddeepnetworkforimagecompressivesensing,in:Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1828–1837

  51. [60]

    Image super-resolutionusingverydeepresidualchannelattentionnetworks, in: Proceedings of the European Conference on Computer Vision (ECCV)

    Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y., 2018. Image super-resolutionusingverydeepresidualchannelattentionnetworks, in: Proceedings of the European Conference on Computer Vision (ECCV). Sriprabha Ramanarayanan et al.:Preprint submitted to ElsevierPage 28 of 28

  52. [170]

    doi:10.1016/j.neuroimage.2017.08.021

  53. [1321]

    doi:10.1109/TMI.2017.2785879

  54. [2017]

    A Deep Cascade of Convolutional Neural Networks for MR ImageReconstruction,in:Proc.Inf.Process.Med.Imaging,pp.647– 658

  55. [2020]

    MAC-ReconNet: A Multiple Acquisition Context based Convolutional Neural Network for MR Image Reconstruction using Dynamic Weight Prediction, in: MIDL, pp. 696–708. URL:http: //proceedings.mlr.press/v121/ramanarayanan20a.html

  56. [2023]

    IEEE Transactions on Medical Imaging 42, 2804–2816

    Multimodal transformer for accelerated mr imaging. IEEE Transactions on Medical Imaging 42, 2804–2816. doi:10.1109/TMI. 2022.3180228

  57. [4148]

    doi:10.1109/ICCV.2019.00424

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.