Pith. sign in

REVIEW 3 major objections 5 minor 25 references

A Frequency-Aware Self-Supervised Learning for Ultra-Wide-Field Image Enhancement

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A two-stage self-supervised framework, FRED+RICE, can deblur and re-illuminate ultra-wide-field retinal images without paired real data, and the improved images raise diabetic retinopathy grading accuracy by 12.3% and F1-score by 15.8%.

desk verdict A plausible UWF enhancement framework with a real gap in validation: the synthetic-blur assumption is unexamined and the evaluation metrics don't carry the weight of the claims. read the letter →

arxiv 2508.19664 v1 pith:BR74373T submitted 2025-08-27 cs.CV

Weicheng Liao , Zan Chen , Jianyang Xie , Yalin Zheng , Yuhui Ma , Yitian Zhao This is my paper · ORCID
classification cs.CV
keywords ultra-wide-fieldimagingretinalimageenhancementself-supervisedlearningfrequency-decoupleddeblurringRetinexilluminationcompensationdiabeticretinopathygradingcolorpreservationwavelettransform
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 ultra-wide-field (UWF) retinal images, which suffer from blur and uneven illumination, can be restored by a self-supervised pipeline that first deblurs in the frequency domain and then corrects illumination with a Retinex-based model. The method, called FRED+RICE, is trained without paired real high-quality UWF images: deblurring uses synthetic blur/clear pairs, and illumination compensation uses reference-free losses. On a 434-image UWF test set, the approach reports the lowest NIQE, BRISQUE and PIQE scores among seven methods, with a large margin over the six baselines. On a separate 902-image diabetic retinopathy grading task, images enhanced by this method improve classification accuracy from 57.13% to 69.49% and F1-score from 48.24% to 64.07% over original images. If true, this means UWF enhancement can be performed without expensive paired data, and the gains carry over to an automated diagnostic task.

What carries the argument

The framework's load-bearing components are the FRED deblurring module and the RICE illumination module. FRED uses average pooling separation to decompose images into high- and low-frequency features, processes each in a separate encoder-decoder stream, and merges them with an asymmetric channel integration (ACI) unit that combines global channel attention (from max and average pooling followed by an MLP) with local details (from two 1x1 convolutions), then feeds through a frequency fusion module. RICE exploits the Retinex identity I = R ⊗ L and estimates a compensation ratio r = L / L' so that the enhanced image is I / r, under the assumption that reflectance stays constant. Its color prese

What would settle it

Collect a small set of real UWF images with corresponding clear references (for example, repeated scans or a different acquisition mode) and compare FRED's output against a network trained on those real pairs; if the synthetic-trained model performs noticeably worse, the degradation assumption fails. Alternatively, measure the blur kernel statistics of real UWF images and compare them to the synthetic kernels used for training.

Watch

Extended reading notes

Core claim

The central claim is that a frequency-aware self-supervised framework can enhance ultra-wide-field retinal images better than generic enhancers. The first stage, FRED, splits the image into high- and low-frequency components using average pooling separation, processes them in dual encoder-decoder streams, and fuses them through an asymmetric channel integration (ACI) unit that mixes global and local information; it is trained on images synthetically blurred by a random degradation operation. The second stage, RICE, applies a Retinex decomposition and estimates an illumination compensation ratio using a color preservation unit (CPU) based on a discrete wavelet transform to prevent color disto

Load-bearing premise

The load-bearing premise is that the random blurring operation applied to high-quality images produces blur that faithfully models the real blur in UWF images, so that a deblurring network trained on these synthetic pairs generalizes to actual UWF scans.

Editorial extensions

If this is right

  • UWF image enhancement becomes feasible without collecting paired real high-quality scans, since the deblurring stage is trained on synthetic blur pairs and the illumination stage uses zero-reference losses.
  • The two-stage sequence—debllur in the frequency domain, then correct illumination with a Retinex model—could transfer to other wide-field imaging modalities where blur and uneven lighting coexist.
  • The color preservation unit offers a concrete way to avoid the color casts that typically plague Retinex-based enhancement on wide-angle retinal images, which may benefit other medical imaging tasks.
  • Enhanced UWF images could serve as a standard preprocessing step for automated diabetic retinopathy screening, potentially improving detection in real-world settings where image quality is inconsistent.
  • If the reported gains replicate, the method provides a clinical tool that improves both human readability and machine-based diagnosis from the same enhanced image.

Reading between the lines

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

  • If the synthetic blur used in training does not match the optical blur of the specific UWF device deployed, the deblurring gains may shrink in practice; a straightforward test is to calibrate the degradation model on a few real blur/clear pairs from the target device.
  • The large jump in DR grading accuracy suggests that some benefit may come from simple intensity normalization or denoising rather than the frequency-decoupled architecture itself; ablating FRED with a generic deblurrer would isolate the source of the gain.
  • The constant-reflectance assumption in RICE could fail near specular reflections or bright artifacts, which are common in UWF periphery; a targeted study on such regions would reveal whether the color preservation unit fully compensates.
  • The enhancement's effect on lesion-level tasks (segmentation, staging sub-classification) is untested; if it improves localization as much as classification, it would strengthen the clinical case.
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

3 major / 5 minor

Summary. The paper proposes a two-stage framework for ultra-wide-field (UWF) retinal image enhancement: a frequency-decoupled deblurring module (FRED) and a Retinex-guided illumination compensation module (RICE). FRED decomposes images into high- and low-frequency streams, uses an Asymmetric Channel Integration (ACI) unit for cross-scale fusion, and is trained on synthetic blur/clear pairs generated by degrading high-quality images. RICE estimates an illumination compensation ratio using a Color Preservation Unit (CPU) based on discrete wavelet transforms. The method is evaluated on a private UWF dataset of 834 images with NIQE, BRISQUE, and PIQE scores, and on a downstream diabetic retinopathy grading task using a ResNet34 classifier. The authors report the best no-reference scores among seven comparison methods and a 12.3% accuracy / 15.8% F1 improvement over original images in DR grading.

Significance. If the results are reliable, this would be a useful contribution: UWF enhancement is an underexplored but clinically important problem, and the proposed architecture is technically well motivated, with a reasonable separation of deblurring and illumination compensation. The downstream DR grading experiment is a valuable attempt to demonstrate clinical utility. However, the current evidence does not yet establish the central claim that the method restores fine details in real UWF images, because the deblurring stage is validated only through synthetic degradation and no-reference metrics. The paper does not provide code, but the method description is sufficiently detailed to be reproducible in principle.

major comments (3)
  1. [§2.1] The FRED module is trained on pairs generated by 'a random blurring degradation operation [14]' applied to 'high-quality images with clear structures', but the paper gives no specification of the blur kernel type, size, spatial variation, or the criterion for 'high-quality'. Real UWF blur is shown in Fig. 1(c) to be prominent in the peripheral retina and is likely spatially varying (e.g., field curvature and laser-scanning optics). A model trained on uniform random blur may not generalize to peripheral blur. Moreover, if the source 'high-quality' images contain residual intrinsic UWF blur, the target is not sharp, so the network may learn only to remove the added synthetic blur. Because the test set (§3) has no sharp ground truth, actual deblurring performance on real UWF images is never directly quantified. This is load-bearing for the claim that FRED restores fine local details; the au
  2. [§3.1, Table 1] The quantitative evaluation relies exclusively on NIQE, BRISQUE, and PIQE, which are no-reference metrics designed for natural images and have not been validated for UWF retinal images. Lower scores do not necessarily imply clinically meaningful improvement or faithful preservation of pathological features. No statistical significance tests, confidence intervals, or multi-run variability are reported; the 'large margin' claim is based on point estimates. The authors should add appropriate statistical analysis (e.g., paired tests with multiple seeds or bootstrapping) and, ideally, a clinical readability evaluation by graders.
  3. [§3.3] The DR grading experiment uses a single ResNet34 with a 7:3 split on 902 images, without reporting confidence intervals or statistical tests. The class distribution is highly imbalanced (444 normal vs. 81 proliferative), and no class-wise metrics or multi-run variance are given. The accuracy/F1 improvements (12.3%/15.8%) could be driven by stochastic variation in the single seed, and the reader cannot assess whether the downstream improvement is significant. Additionally, the paper does not state whether the split is at the patient level; if multiple images from the same eye or patient appear in both train and test sets, results are optimistic. Please provide details and significance testing.
minor comments (5)
  1. [General] There are typographical issues: missing spaces in the abstract and introduction, e.g., 'thepreservationoffineandbroaderstructuraldetails.Inaddition,acolor'.
  2. [§2.1, Eq. (1)] Notation is inconsistent: 'M P' appears with a space in the definition of f^m_s, while 'MP' is used elsewhere. Please unify notation.
  3. [§2.2, Eq. (4)] The notation H^2_1 and H^2_5 is not defined clearly; it is inferred from the text to mean two successive 1×1 and 5×5 convolutional layers, but this should be explicit in the equation or immediately after.
  4. [§3.2] The ablation study reports relative percentage changes in NIQE/BRISQUE/PIQE without confidence intervals or significance tests, so the reader cannot determine whether the differences between the ablations and the full model are meaningful. This echoes the statistical concern in the main evaluation.
  5. [Introduction] The phrase 'peripheral distortions obtained by unsupervised approach' is grammatically unclear and should be rewritten. Also, the term 'self-supervised' is used for the overall method, but the deblurring stage is trained with explicit paired synthetic supervision; the terminology could be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the claimed enhancement and grading improvements are empirical, trained on synthetic blur pairs and evaluated on held-out UWF images; no output quantity is equivalent to a fitted input by construction.

full rationale

The paper's derivation chain is not circular. FRED (§2.1) is trained on paired blur/clear images produced by applying a random blurring degradation operation [14] to high-quality images. The network learns a mapping from synthesized degradations; at test time it is applied to real UWF images and assessed with no-reference metrics and a downstream DR grading task. No test-set quantity is used to fit the network parameters, and no loss term trivially guarantees the reported NIQE/BRISQUE/PIQE improvements. RICE (§2.2) uses a standard Retinex decomposition I=R⊗L and estimates a compensation ratio r = L ⊘ L' with a learned network; the enhancement I' = I ⊘ r is not definitionally equal to a fitted target. The DR grading experiment uses an additional independent dataset (902 UWF images) with a classifier trained on original images and tested on enhanced images, so the 12.3% ACC and 15.8% F1 gains are external validation, not self-fulfilling. The label 'self-supervised' is technically inaccurate because the FRED training uses paired supervision on synthetic data, but this is a terminology issue, not circularity. The only substantive caveat is that synthetic blur may not perfectly match real UWF blur distribution; this is a generalization/validity concern, not a circularity. Self-citations ([11], background) are not load-bearing to the method's architecture or results. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method adds two hand-designed modules (ACI and CPU) and three manually set loss weights. The main external assumptions are the fidelity of synthetic blur to real UWF blur and the validity of no-reference metrics. No new physical entities are introduced.

free parameters (3)
  • beta (loss weight for MSFR loss) = 0.1
    Set manually in Section 3 without sensitivity analysis; affects Eq. 2.
  • gamma (loss weight for perceptual loss) = 0.01
    Set manually in Section 3 without sensitivity analysis; affects Eq. 2.
  • alpha (loss weight for fidelity loss) = 1.5
    Set manually in Section 3 without sensitivity analysis; affects Eq. 5.
assumptions (4)
  • domain assumption Synthetic blur from random degradation approximates real UWF blur
    Section 2.1: paired blur/clear samples are generated by applying random blurring to high-quality images; the model's deblurring ability on real UWF images depends on this.
  • domain assumption No-reference metrics NIQE, BRISQUE, PIQE are valid for UWF retinal image quality
    Section 3.1: these metrics are used as primary quantitative evaluation without validation on UWF.
  • domain assumption Retinex reflectance constancy (R' = R) after enhancement
    Section 2.2: assumes reflectance remains unchanged during illumination compensation, which may not hold in peripheral UWF regions.
  • standard math Discrete wavelet transform with the four filters in Eq. 3 is a standard, invertible transform
    Section 2.2: used for frequency separation in CPU; standard signal processing result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Frequency-Aware Self-Supervised Learning for Ultra-Wide-Field Image Enhancement." pith.science (2026). https://pith.science/paper/BR74373T

@misc{pith2026250819664,
  author       = {Pith},
  title        = {Pith review of: A Frequency-Aware Self-Supervised Learning for Ultra-Wide-Field Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BR74373T}},
  note         = {Machine review of arXiv:2508.19664}
}
read the original abstract

Ultra-Wide-Field (UWF) retinal imaging has revolutionized retinal diagnostics by providing a comprehensive view of the retina. However, it often suffers from quality-degrading factors such as blurring and uneven illumination, which obscure fine details and mask pathological information. While numerous retinal image enhancement methods have been proposed for other fundus imageries, they often fail to address the unique requirements in UWF, particularly the need to preserve pathological details. In this paper, we propose a novel frequency-aware self-supervised learning method for UWF image enhancement. It incorporates frequency-decoupled image deblurring and Retinex-guided illumination compensation modules. An asymmetric channel integration operation is introduced in the former module, so as to combine global and local views by leveraging high- and low-frequency information, ensuring the preservation of fine and broader structural details. In addition, a color preservation unit is proposed in the latter Retinex-based module, to provide multi-scale spatial and frequency information, enabling accurate illumination estimation and correction. Experimental results demonstrate that the proposed work not only enhances visualization quality but also improves disease diagnosis performance by restoring and correcting fine local details and uneven intensity. To the best of our knowledge, this work is the first attempt for UWF image enhancement, offering a robust and clinically valuable tool for improving retinal disease management.

Figures

Figures reproduced from arXiv: 2508.19664 by the authors.

Figure 1
Figure 1. Illustrations of UWF imaging. (a-b) A color fundus image and an UWF [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework and its main components. (a) Overall [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The visualization results by using different methods. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 23 canonical work pages

  1. [14]

    In: International Workshop on Ophthalmic Medical Image Analysis

    Pérez, A.D., Perdomo, O., Rios, H., Rodríguez, F., González, F.A.: A conditional generativeadversarialnetwork-basedmethodforeyefundusimagequalityenhance- ment. In: International Workshop on Ophthalmic Medical Image Analysis. pp. 185–194. Springer (2020)

    work page 2020

  2. [1]

    Automation in Construction 162, 105404 (2024)

    Chen, X., Yu, Y.: An unsupervised low-light image enhancement method for im- proving v-slam localization in uneven low-light construction sites. Automation in Construction 162, 105404 (2024)

    work page 2024

  3. [2]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine ap- proach in single image deblurring. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4641–4650 (October 2021)

    work page 2021

  4. [3]

    IEEE Journal of Biomedical and Health Informatics26(9), 4645–4655 (2022)

    Deng, Z., Cai, Y., Chen, L., Gong, Z., Bao, Q., Yao, X., Fang, D., Yang, W., Zhang, S., Ma, L.: Rformer: Transformer-based generative adversarial network for real fundus image restoration on a new clinical benchmark. IEEE Journal of Biomedical and Health Informatics26(9), 4645–4655 (2022)

    work page 2022

  5. [4]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Dong, J., Pan, J., Yang, Z., Tang, J.: Multi-scale residual low-pass filter network for image deblurring. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12345–12354 (2023)

    work page 2023

  6. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)

    Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)

    work page 2020

  7. [6]

    PLoS One18(3), e0282416 (2023)

    Lee, K.G., Song, S.J., Lee, S., Yu, H.G., Kim, D.I., Lee, K.M.: A deep learning- based framework for retinal fundus image enhancement. PLoS One18(3), e0282416 (2023)

    work page 2023

  8. [7]

    IEEE Transactions on Pattern Analysis and Machine Intel- ligence 44(8), 4225–4238 (2022) 10 W

    Li, C., Guo, C., Loy, C.C.: Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intel- ligence 44(8), 4225–4238 (2022) 10 W. Liao et al

    work page 2022

Show all 25 references
  1. [8]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Liu, H., Li, H., Fu, H., Xiao, R., Gao, Y., Hu, Y., Liu, J.: Degradation-invariant enhancement of fundus images via pyramid constraint network. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 507–516. Springer (2022)

    2022

  2. [9]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Liu, R., Ma, L., Zhang, J., Fan, X., Luo, Z.: Retinex-inspired unrolling with cooper- ative prior architecture search for low-light image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10561–10570 (June 2021)

    2021

  3. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5637–5646 (June 2022)

    2022

  4. [11]

    In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part II 23

    Ma, Y., Liu, Y., Cheng, J., Zheng, Y., Ghahremani, M., Chen, H., Liu, J., Zhao, Y.: Cycle structure and illumination constrained gan for medical image enhancement. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, ...

    2020

  5. [12]

    IEEE Transactions on image processing 21(12), 4695–4708 (2012)

    Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695–4708 (2012)

    2012

  6. [13]

    completely blind

    Mittal, A., Soundararajan, R., Bovik, A.C.: Making a “completely blind” image quality analyzer. IEEE Signal Processing Letters20(3), 209–212 (2012)

    2012

  7. [15]

    IEEE Transactions on Medical Imaging40(3), 996–1006 (2021)

    Shen, Z., Fu, H., Shen, J., Shao, L.: Modeling and enhancing low-quality retinal fundus images. IEEE Transactions on Medical Imaging40(3), 996–1006 (2021)

    2021

  8. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Shi, Y., Liu, D., Zhang, L., Tian, Y., Xia, X., Fu, X.: Zero-ig: Zero-shot illumination-guided joint denoising and adaptive enhancement for low-light images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3015–3024 (2024)

    2024

  9. [17]

    In: 2015 twenty first national conference on communications (NCC)

    Venkatanath, N., Praneeth, D., Bh, M.C., Channappayya, S.S., Medasani, S.S.: Blind image quality evaluation using perception based features. In: 2015 twenty first national conference on communications (NCC). pp. 1–6. IEEE (2015)

    2015

  10. [18]

    In: 2008 International Conference on Wavelet Analysis and Pattern Recognition

    Wang, W., Li, B., Zheng, J., Xian, S., Wang, J.: A fast multi-scale retinex algo- rithm for color image enhancement. In: 2008 International Conference on Wavelet Analysis and Pattern Recognition. vol. 1, pp. 80–85 (2008)

    2008

  11. [19]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, W., Yang, H., Fu, J., Liu, J.: Zero-reference low-light enhancement via phys- ical quadruple priors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 26057–26066 (2024)

    2024

  12. [20]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wu, Y., Pan, C., Wang, G., Yang, Y., Wei, J., Li, C., Shen, H.T.: Learning semantic-aware knowledge guidance for low-light image enhancement. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1662–1671 (2023)

    2023

  13. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Xu, X., Wang, R., Lu, J.: Low-light image enhancement via structure modeling and guidance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9893–9903 (2023)

    2023

  14. [22]

    arXiv preprint arXiv:2404.14132 (2024) Frequency-Aware UWF Image Enhancement 11

    Yang,K.,Hu,T.,Dai,K.,Chen,G.,Cao,Y.,Dong,W.,Wu,P.,Zhang,Y.,Yan,Q.: Crnet: A detail-preserving network for unified image restoration and enhancement task. arXiv preprint arXiv:2404.14132 (2024) Frequency-Aware UWF Image Enhancement 11

    2024 arXiv

  15. [23]

    Expert Systems with Applications 244, 123005 (2024)

    Zhang, B., Sun, J., Sun, F., Wang, F., Zhu, B.: Image deblurring method based on self-attention and residual wavelet transform. Expert Systems with Applications 244, 123005 (2024)

    2024

  16. [24]

    International Journal of Computer Vision 130(9), 2103–2130 (2022)

    Zhang, K., Ren, W., Luo, W., Lai, W.S., Stenger, B., Yang, M.H., Li, H.: Deep image deblurring: A survey. International Journal of Computer Vision 130(9), 2103–2130 (2022)

    2022

  17. [25]

    In: Graphics gems IV

    Zuiderveld, K.: Contrast limited adaptive histogram equalization. In: Graphics gems IV. pp. 474–485. Academic Press Professional, Inc. (1994)

    1994

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.