{"id":"42cc0ae8-1b4a-4d11-b42d-4b21c1233f11","arxiv_id":"2508.19664","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A frequency-aware two-module network (FRED+RICE) enhances ultra-wide-field retinal images, improving no-reference quality scores and boosting DR grading accuracy.","lead":"This paper presents a two-stage neural network that first deblurs ultra-wide-field retinal images and then corrects uneven illumination, using frequency-aware modules. The authors report that the enhanced images improve a diabetic retinopathy grading network's accuracy by 12.3 percentage points over the original images.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FRED's deblurring is trained on synthetic uniform blur but real UWF blur is spatially varying; if the synthetic degradation doesn't match, the central claim of restored fine details is unvalidated.","rationale":"I agree with the reader's identification of the synthetic blur degradation as the key unvalidated assumption. The full pipeline evaluation mixes the effects of FRED and RICE; the grading improvement could be driven primarily by RICE's illumination correction, while FRED's deblurring might contribute only by introducing a sharpening artifact that helps the classifier but does not faithfully restore real UWF blur. The paper provides no direct validation of deblurring on real images because no ground truth is available. The proposed test would isolate whether the trained FRED generalizes to a more realistic spatially varying blur. Other issues, such as the reliability of NIQE/BRISQUE/PIQE for UWF and the absence of significance testing in the grading experiment, are also relevant but they affect the strength of the overall claims rather than the core mechanism of the proposed method. Since the reader's verdict already requires additional validation and code/data release, my analysis does not change the verdict: the paper remains CONDITIONAL pending evidence that the synthetic blur setup captures real UWF blur.","tokens_in":7698,"tokens_out":7468,"duration_ms":84661,"concrete_test":"Construct a validation set by applying a realistic UWF blur model (e.g., Gaussian blur with standard deviation increasing with distance from the image center, simulating peripheral aberration) to a held-out set of high-quality UWF images. Degrade each image with this spatially varying blur and also with the uniform random blur used in training. Run the trained FRED model on both degraded versions and compare deblurred outputs to the originals using PSNR/SSIM. If performance on the spatially varying blur is substantially worse (e.g., PSNR drop >2 dB) or the output exhibits residual peripheral blur/artifacts, the training degradation is not representative and the FRED deblurring on real UWF is unvalidated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The FRED module (§2.1) is trained on pairs generated by 'a random blurring degradation operation [14]' applied to 'high-quality images with clear structures'. The paper gives no specification of the blur kernel (type, size, spatial variation) nor how 'high-quality' is determined. Real UWF blur, as shown in Fig. 1(c), is especially prominent in the peripheral retina and is likely to be spatially varying (e.g., field-curvature and laser-scanning induced). A model trained on uniform random blur may not deblur peripheral regions correctly. Moreover, if the 'clear' source images themselves contain residual UWF blur, the target for deblurring is not sharp, so the network learns only to remove the added synthetic blur, leaving intrinsic blur untouched. Because the test set (§3) has no ground-truth sharp images, the deblurring effectiveness on real UWF is never directly quantified; the only evidence is indirect (improved no-reference scores and DR grading). If the synthetic-blur assumption fails, the claimed 'restoration of fine local details' and the contribution of FRED to the grading improvement are not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":1643,"tokens_out":1516,"duration_ms":48660,"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":[{"comment":"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","section":"§2.1"},{"comment":"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.","section":"§3.1, Table 1"},{"comment":"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.","section":"§3.3"}],"minor_comments":[{"comment":"There are typographical issues: missing spaces in the abstract and introduction, e.g., 'thepreservationoffineandbroaderstructuraldetails.Inaddition,acolor'.","section":"General"},{"comment":"Notation is inconsistent: 'M P' appears with a space in the definition of f^m_s, while 'MP' is used elsewhere. Please unify notation.","section":"§2.1, Eq. (1)"},{"comment":"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.","section":"§2.2, Eq. (4)"},{"comment":"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.","section":"§3.2"},{"comment":"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.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a workmanlike application paper that pairs a frequency-decoupled deblurring stage (FRED) with a Retinex-based illumination compensation stage (RICE) for ultra-wide-field retinal images. The claimed first application to UWF is credible, and the architecture is a sensible combination of existing building blocks (CRNet's average-pooling frequency separation, SCI's Retinex losses, wavelet filters for color preservation). The paper is clearly written, the ablations are standard, and the downstream DR grading experiment is a good idea. So there is real substance here, not a toy.\n\nThe soft spots are all in evaluation. The biggest is the synthetic blur assumption. The FRED module is trained on pairs produced by applying a random blur to 'high-quality' images, but the paper never specifies the blur kernel, its size, or whether it is spatially varying. Real UWF blur is peripheral and spatially varying. If the synthetic degradation doesn't match, the model only learns to remove the added blur, not the intrinsic UWF blur. The stress-test note is exactly right: there is no ground truth on the test set, so deblurring performance on real images is never directly measured.\n\nThe quality metrics are the second problem. NIQE, BRISQUE, and PIQE are no-reference metrics designed for natural images; their validity on UWF retinal images is unproven, and 'large margin' claims based on them are shaky. The DR grading experiment is a single ResNet34 run without significance tests or confidence intervals, so the 12.3% accuracy gain could be noise. No code or data is provided.\n\nNone of this is fatal to the method. The components are plausible, the ablations show each part helps on the chosen metrics, and the clinical motivation is real. The gap is validation, not concept. A serious editor should send this to peer review and let the referees ask for: the blur model specification and ideally a test against real paired or clinical data; significance testing on both metrics and grading; and code/data release or a supplementary with per-image results. Until then, the claimed superiority is unproven.\n\nThis paper is for researchers working on retinal image enhancement or on the gap between synthetic training and real clinical data. I'd bring it to a reading group; I would not cite it in my own work yet. It deserves a serious referee, not a desk reject.","headline":"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.","tokens_in":8453,"tokens_out":3748,"would_cite":false,"duration_ms":36489,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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%.","keywords":["ultra-wide-field imaging","retinal image enhancement","self-supervised learning","frequency-decoupled deblurring","Retinex illumination compensation","diabetic retinopathy grading","color preservation","wavelet transform"],"falsifier":"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.","tokens_in":7646,"feed_emoji":"👁️","tokens_out":5805,"duration_ms":50768,"temperature":0.7,"pith_summary":"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.","feed_headline":"UWF eye-image enhancer raises DR grading accuracy by 12.3%","feed_subtitle":"Two-stage self-supervised model deblurs and re-lights wide-field retinal scans, improving both clarity and diagnosis.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the random blurring degradation operation used to synthesize paired blur/clear training samples for FRED.","marker":"[14]"},{"why":"Provides the average pooling separation (APS) and frequency fusion module (FFM) that FRED is built on.","marker":"[22]"},{"why":"Supplies the Retinex theory and smoothness loss used in RICE for illumination decomposition and estimation.","marker":"[10]"},{"why":"Provides the exposure control loss used in RICE's joint loss function.","marker":"[5]"},{"why":"Defines the multi-scale frequency reconstruction loss used to train FRED.","marker":"[2]"},{"why":"Provides the perceptual loss used in FRED's training objective.","marker":"[24]"},{"why":"Defines NIQE, the no-reference quality metric used to evaluate enhancement performance.","marker":"[13]"},{"why":"Defines BRISQUE, the no-reference quality metric used to evaluate enhancement performance.","marker":"[12]"},{"why":"Defines PIQE, the no-reference quality metric used to evaluate enhancement performance.","marker":"[17]"}],"fun_headline_variants":["First UWF enhancer uses frequency-aware learning to lift clarity and DR grading","Self-supervised model deblurs and re-lights retinal scans, boosting DR diagnosis by 12.3%","Frequency-decoupled AI restores fine detail in ultra-wide-field eye images","New method combines deblurring and Retinex to enhance UWF imaging for better clinical deci"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["First UWF enhancer uses frequency-aware learning to lift clarity and DR grading","Self-supervised model deblurs and re-lights retinal scans, boosting DR diagnosis by 12.3%","Frequency-decoupled AI restores fine detail in ultra-wide-field eye images","New method combines deblurring and Retinex to enhance UWF imaging for better clinical decisions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1426,"prompt_tokens":758,"completion_tokens":668,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":573}},"tokens_in":502,"tokens_out":668,"duration_ms":6809,"temperature":1.0,"reasoning_tokens":573,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:33:46.772297+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"In: International Workshop on Ophthalmic Medical Image Analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the random blurring degradation operation used to synthesize paired blur/clear training samples for FRED."},{"cited_title":"CRNet: A Detail-Preserving Network for Unified Image Restoration and Enhancement Task","cited_arxiv_id":"2404.14132","evidence_quote":"Provides the average pooling separation (APS) and frequency fusion module (FFM) that FRED is built on."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Supplies the Retinex theory and smoothness loss used in RICE for illumination decomposition and estimation."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)","cited_arxiv_id":null,"evidence_quote":"Provides the exposure control loss used in RICE's joint loss function."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)","cited_arxiv_id":null,"evidence_quote":"Defines the multi-scale frequency reconstruction loss used to train FRED."},{"cited_title":"International Journal of Computer Vision 130(9), 2103–2130 (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the perceptual loss used in FRED's training objective."},{"cited_title":"IEEE Transactions on image processing 21(12), 4695–4708 (2012)","cited_arxiv_id":null,"evidence_quote":"Defines BRISQUE, the no-reference quality metric used to evaluate enhancement performance."},{"cited_title":"In: 2015 twenty first national conference on communications (NCC)","cited_arxiv_id":null,"evidence_quote":"Defines PIQE, the no-reference quality metric used to evaluate enhancement performance."}],"review_version":1}