acceptodds
Withdrawn from ICLR 2027

Coupled Spatial–Spectral Learning for Generative MRI Reconstruction with Enhanced Diagnostic Separability

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accelerated Magnetic Resonance Imaging (MRI) reconstruction is fundamentally an ill-posed inverse problem where aggressive undersampling of the frequency domain (k-space) leads to severe aliasing and the loss of high-frequency structural details. While deep generative adversarial networks (GANs) have established new benchmarks for this task, they frequently struggle with frequency-domain consistency, yielding reconstructions with over-smoothed textures and blurred semantic boundaries. To address this and improve diagnostic separability, we introduce AUSGAN, a novel generative framework that couples spatial attention with spectral regularization to enforce structural fidelity across dual domains. Specifically, the generator integrates soft attention mechanisms to dynamically gate high-frequency features along semantic edges, while the discriminator employs an azimuthal integration module to explicitly penalize distributional shifts in the polar spectral profile. This coupled spatial-spectral synergy prevents the hallucination of isotropic noise while strictly constraining high-frequency recovery to valid anatomical regions. We benchmark AUSGAN against state-of-the-art reconstruction baselines under highly aggressive undersampling rates (up to 10x acceleration) across three diverse clinical cohorts: the ISBR dataset and multi-modal datasets for Brain Metastasis and Liver Cirrhosis. Going beyond standard pixel-wise metrics, we introduce a boundary-aware evaluation framework leveraging the Bhattacharyya distance to quantitatively measure the feature separability of the reconstructed outputs. Extensive experiments demonstrate that AUSGAN establishes a new state-of-the-art, delivering superior perceptual quality, robust artifact suppression, and significantly enhanced semantic contrast at critical tissue interfaces.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Accelerated MRI Reconstruction, Generative Adversarial Networks (GANs), Spatial-Spectral Learning, Spectral Regularization, Spatial Attention, Diagnostic Separability

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_2ucsOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes AUSGAN, a dual-domain GAN framework for accelerated MRI reconstruction, combining k-space and image-space generators, attention-gated feature refinement, and a spectral discriminator. It further introduces a boundary-aware metric based on Bhattacharyya distance to assess tissue separability. Experiments are conducted on three MRI cohorts under different sampling rates, with comparisons against several reconstruction baselines.

Critical strengths
  1. The paper attempts to evaluate MRI reconstruction beyond conventional PSNR and SSIM by considering tissue-boundary separability, which is potentially relevant to downstream diagnosis.

  2. The dual-domain reconstruction framework is reasonably structured, and the experiments include multiple datasets and component ablations.

Critical weaknesses
  1. The proposed framework largely combines established techniques, including dual-domain reconstruction, attention-gated U-Nets, adversarial learning, and frequency-domain constraints. The paper does not clearly establish a sufficiently distinct methodological contribution beyond integrating existing components.

  2. AUSGAN reportedly achieves SSIM values of 0.97–0.99 and PSNR exceeding 40 dB even at 10% k-space sampling, substantially outperforming competing methods. These unusually strong results require careful verification, particularly given inconsistencies in numerical reporting and unusually large standard deviations in some tables. The absence of key quantitative results from the main text further limits reproducibility and confidence in the findings.

  3. Although the paper claims that spectral discrimination suppresses hallucinated structures and improves diagnostic fidelity, the ablation results show only marginal changes after removing the spectral component. No direct evaluation of hallucination, anatomical preservation, or frequency-specific reconstruction errors is provided. Consequently, the claimed mechanism is not convincingly supported.

  4. The proposed Bhattacharyya-distance metric lacks clinical validation and sometimes assigns higher separability to reconstructed images than to the original reference images. This raises concerns that the metric may reward artificial contrast enhancement rather than faithful anatomical reconstruction. Without radiologist assessment or established downstream diagnostic evaluation, the claimed clinical benefits remain speculative.

  5. The paper contains broken figure references, inconsistent numerical formatting, and unclear descriptions of multimodal inputs. Moreover, the argument regarding computational efficiency is unsupported by inference-time or parameter-count comparisons.

Questions for authors
  1. How can the unusually high PSNR and SSIM at 10% sampling be explained?
  2. Why does the proposed separability metric sometimes favor reconstructed images over the original references?
  3. How does AUSGAN differ fundamentally from existing dual-domain GAN architectures, beyond combining established modules?
  4. Can the authors clarify the multimodal experimental settings and provide computational efficiency comparisons?
Reviewer_7RUJOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes AUSGAN, a dual-domain GAN for undersampled MRI reconstruction. It combines adjacent-slice k-space completion, attention-based image refinement, and a discriminator using spatial features and azimuthally averaged spectral information. Evaluation on brain and liver MRI cohorts includes reconstruction metrics, component ablations, and a boundary-based Bhattacharyya distance intended to assess tissue separability.

Critical strengths
  1. AUSGAN combines adjacent-slice information, dual-domain reconstruction, and spatial–spectral adversarial supervision in a coherent framework. Preserving acquired k-space samples in the intermediate reconstruction provides a useful measurement constraint, while the complementary spatial and spectral components offer a technically motivated approach to recovering structural detail under undersampling.
  2. The evaluation includes both healthy anatomy and pathological cohorts, with additional analysis of clinically relevant tissue interfaces. The boundary-focused evaluation and component ablations provide useful perspectives on reconstruction behavior that global image metrics alone may overlook.
Critical weaknesses
  1. The comparative results contain substantial internal inconsistencies that undermine the evidence for the claimed performance gains. Complete CS-MRI and AdaDiff PSNR rows, including uncertainties, are repeated across the ISBR and liver tables, while the liver table reports CS-MRI outperforming AUSGAN in several settings despite marking AUSGAN as best. These discrepancies directly affect the central claim of consistent superiority. Corrected tables and reproducible evaluation records are needed to determine whether the problems are limited to reporting or also affect the underlying results.
  2. The claims of improved diagnostic fidelity and hallucination prevention are not sufficiently supported. Bhattacharyya distance measures tissue-intensity separation and can increase through contrast alteration without improving anatomical fidelity. AUSGAN’s reported values consistently exceed those of the reference images, but the paper provides no direct lesion-level or diagnostic evaluation to distinguish faithful recovery from contrast enhancement. Moreover, acquired-sample consistency is explicitly enforced only before the final image refinement. The current evidence therefore does not establish that clinically relevant structures are preserved without additions, distortions, or omissions.
Questions for authors
  1. Can the authors reconcile the repeated baseline rows and inconsistent rankings in Tables 2 and 4, and provide corrected results together with the metric implementations and patient-level aggregation procedure? In particular, is the SSIM formula in Equation (2) a typographical error or the formula used in evaluation?
  2. How do the authors distinguish faithful anatomical recovery from contrast alteration when Bhattacharyya distances exceed those of the reference images? Could this be assessed through lesion-level fidelity or an independent diagnostic task or reader evaluation?
  3. Is acquired k-space consistency enforced after the final image-domain generator? Please clarify how the final output is related to the acquired measurements and report its residual at the sampled k-space locations.
Reviewer_1dFQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an accelerated MRI reconstruction based on generative adversarial networks, named AUSGAN. The proposed AUSGAN couples spatial attention with spectral regularization, and incorporates image and k-space domains. Experiments are performed against multiple datasets and over several competing methods.

Critical strengths

The authors address an important problem in accelerated MRI reconstruction. The paper is well written, it is clear, and easy to follow. The main diagrams effectively explain the proposed methodology and help understand the overall approach.

Critical weaknesses
  1. The novelty of the paper is limited. The use of generative adversarial networks for MR image reconstruction has been extensively investigated in prior work. Similarly, hybrid image and k-space reconstruction methods are established, as well as k-space and image-domain loss functions. Based on these, I believe the contribution of the paper is not sufficiently clear.

  2. In clinical practice, MR imaging protocols use multi-coil acquisitions. However, the datasets in this study appear to be single-coil, limiting the relevance and reliability of the experiments. Also, the authors denote that images were pre-processed with skull stripping and resizing. These operations change the original imaging characteristics and introduce inconsistencies. While skull stripping can be a part of other MRI applications such as image synthesis or segmentations, it's not applicable to do it in reconstruction.

  3. The competing methods do not accurately reflect the current state-of-the-art. Most competing methods are from earlier years, and comparison methods against recent diffusion based approaches is limited. In order to demonstrate the efficacy of the proposed method, it must be shown against a more comprehensive and recent set of competing methods.

  4. In several parts of the results section, the figure reference appears to be missing (Figure ??).

  5. Figure 5: The results shown in this figure indicate that the proposed AUSGAN introduce hallucinations, whereas most baselines such as CS-MRI and Zero-filled more accurately reconstruct the target image. Also, why is the result shown for ADMMNet so noisy?

  6. Results in Figure 5 present at an acceleration ratio of R=2x, which is a relatively modest level of undersampling. Also, given that the experiments use single-coil magnitude images, for which conjugate symmetry applies, the reconstruction task is straightforward. This is suggested by the strong performance of the zero-filled approach. Consequently, the added value with this demonstration is unclear. Can the authors provide results at higher acceleration ratios to demonstrate the effectiveness of their approach?

  7. Overall, the experimentation and results are limited. Most results are shown for relatively low acceleration ratios. Results are not supported with visual findings. Looking at the only Figure in the main manuscript, the performance gains by the proposed method are unclear.

  8. The quantitative measurements are reported based on PSNR, SSIM, and Bhattacharyya distance. However, no radiological evaluations or downstream analyses are performed.

Questions for authors
  1. Could you please better explain the main novelty of the paper?

  2. How does the reconstruction performance look when it reflects a more realistic scenario of multi-coil image reconstruction with no additional preprocessing like skull stripping?

  3. How does the proposed method compare against more recent state-of-the-art methods based on diffusion models or unrolled networks?

  4. How does the reconstruction performance visually and quantitatively look on higher acceleration ratios and different anatomies?

  5. Are there any clinical evaluations or downstream analyses on top of standard PSNR measurements etc.?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.