{"id":"3487d0cb-6b75-41f9-9bd8-7d54999afad5","arxiv_id":"2411.14663","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"An attention-augmented hierarchical VQ-VAE (BrightVAE) claims state-of-the-art luminosity enhancement on the Endo4IE endoscopic image dataset, but the evidence is not independently verifiable from the paper alone.","lead":"BrightVAE, an attention-enhanced hierarchical VQ-VAE, is presented to brighten underexposed endoscopic images, with claimed top scores in PSNR, SSIM, and LPIPS on the Endo4IE benchmark. If verified, it would give clinicians and downstream algorithms a stronger preprocessing step, but the release lacks code, error bars, and a clear validation split.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Loss-function ablation in §4.D.2 appears to select LSimilarity on the Endo4IE test set; without a held-out validation split, the reported SOTA margins in Table 1 are not unbiased.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the LSimilarity ablation in Section 4.D.2 appears to have been evaluated on the Endo4IE test set, with no held-out validation split stated. This is the most critical issue because the paper's central assertion is superiority on Endo4IE metrics, and selecting the loss by those same metrics on the test set invalidates the unbiasedness of the comparison. The concern is concrete and testable: a simple re-ablation with a validation split and seeded repeats would determine whether the claimed margins hold. Other issues, such as missing code, lack of error bars, and the synthetic nature of Endo4IE, are secondary; they affect generalizability and reproducibility but do not directly threaten the internal validity of the benchmark comparison as strongly as test-set selection does. Because the reader's verdict CONDITIONAL already captures this, no verdict change is needed; our stress test agrees and reinforces the condition.","tokens_in":10680,"tokens_out":2482,"duration_ms":26617,"concrete_test":"Re-run the loss ablation using a held-out validation split: randomly reserve 20% of the 690 training image pairs as validation, train BrightVAE with each of the eight LSimilarity candidates on the remaining 80%, and select the best candidate by validation PSNR/SSIM. Retrain the final model with that fixed loss and evaluate once on the 266 test images. Compare test PSNR/SSIM/LPIPS against Table 1 and against Mou et al., reporting mean ± std over at least three seeds. If the margin falls below the reported ~3.25 dB or becomes statistically insignificant, the SOTA claim is a test-set selection artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BrightVAE achieves top PSNR, SSIM, and LPIPS on Endo4IE (Section 4.C, Table 1). A key step is the selection of LSimilarity in Section 4.D.2, where eight loss candidates are compared and the SSI loss is chosen because it yields the best PSNR/SSIM. The paper never states that this ablation was performed on a held-out validation set, and Section 4.B describes only a training set (690 pairs) and a test set (266 images). If the test set was used to pick the loss, then the final numbers in Table 1 are not an independent evaluation: the model and its loss have been selected to maximize the exact metrics on the exact benchmark. The reported advantage of ~3.25 dB PSNR over Mou et al. and ~3.34 dB over VQ-VAE could shrink or disappear when the loss is fixed on a validation split before touching the test set. No error bars, standard deviations, or multiple-seed results are reported, so it is also unclear whether even the remaining differences are statistically significant. This directly undermines the paper's strongest claim, making the SOTA assertion unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes BrightVAE, a hierarchical VQ-VAE architecture for low-light endoscopic image enhancement. The main architectural additions are an attention-based encoder (Attencoder), an attention-based quantization module (Attenquant), dual global/local receptive fields, and skip connections between encoder and decoder. The training loss combines an MSE reconstruction term, a VQ latent loss, and a similarity loss that is selected by an ablation over eight candidates. Experiments on the Endo4IE dataset compare BrightVAE with 14 baselines plus VQ-VAE and report that it achieves the highest PSNR, SSIM, and LPIPS, with claimed gains of about 3.25 dB PSNR over Mou et al. and 3.34 dB over VQ-VAE. The paper also includes component and loss-function ablations.","tokens_in":10939,"tokens_out":5136,"duration_ms":47055,"significance":"If the empirical claims hold, BrightVAE would be a useful new state of the art on the public Endo4IE benchmark, and the two attention-based modules are simple enough to be adapted by other researchers. The paper's strengths are that it targets a concrete medical imaging task, uses a public benchmark, compares against many recent methods, and includes ablations. The central SOTA claim is falsifiable on a public dataset. However, the contribution is primarily empirical, and the current evidence is incomplete: no code or pretrained weights are provided, the comparison protocol is underspecified, and no variance estimates are reported. In particular, the loss-function selection may have been performed on the test set, which would make the reported SOTA margins unverified. With additional details and a clean evaluation protocol, the work could be a solid benchmark contribution.","major_comments":[{"comment":"The selection of LSimilarity is load-bearing for the paper's central claim, but the paper never states which data were used for this ablation. Section 4.B describes only a training set (690 pairs) and a test set (266 images); no validation split is mentioned. If the eight loss candidates were compared on the same Endo4IE test images subsequently used in Table 1, then the model has been chosen by optimizing the exact metrics on the test set, and the reported SOTA margins in Section 4.C are not an independent evaluation. Please state explicitly the split used for Table 3 and, if necessary, redo the loss selection on a held-out validation split before computing final test numbers.","section":"Section 4.D.2 (Table 3)"},{"comment":"The protocol for comparing with the 14 baselines is not specified. It is unclear whether each baseline was retrained on the Endo4IE training split, used with publicly released checkpoints, or evaluated by the authors, and whether any pre/post-processing was applied. Additionally, Table 1 is embedded as an image rather than as a text table; the surrounding text gives only the improvements over two methods, not the full numerical values for all methods. Without the protocol and the full numeric table, the claim that BrightVAE 'surpasses all other models' cannot be checked or reproduced.","section":"Section 4.C (Table 1)"},{"comment":"No error bars, standard deviations, or multiple-seed results are reported, and no statistical test is applied. The claimed advantages over the nearest competitor are small in some metrics (0.045 SSIM and 0.014 LPIPS), so it is possible that the ranking is within run-to-run variation. Please report results over at least three training runs for the proposed method (and ideally for baselines) or otherwise characterize the variability, so the reader can judge whether the stated margins are meaningful.","section":"Section 4.C"}],"minor_comments":[{"comment":"The equations for LRest and LLatent are missing from the text; only prose descriptions and placeholders appear. Please include the actual formulas.","section":"Section 3.E"},{"comment":"Tables 1, 2, and 3 appear only as figures; the numeric values are not available to the reader. Please typeset them as real tables.","section":"Tables 1-3"},{"comment":"The paper uses both 'SSI loss' and 'SSIM loss' for the same quantity; choose one term and use it consistently.","section":"Sections 4.D.2 and 4.E"},{"comment":"There is a redundant phrase: 'for local branch operations aimed at local branch.' Please rephrase.","section":"Section 3.A"},{"comment":"Specific hyperparameters (codebook size, attention heads, embedding dimension, learning-rate schedule values) are not given; these details are needed to reproduce the architecture.","section":"Section 4.A"},{"comment":"Reference [15] is mentioned in Related Work but does not appear among the methods compared in Section 4.C; please clarify its status.","section":"Related Work and Section 4.C"},{"comment":"The phrase 'significant advancements' is not supported by statistical testing; please either soften it or add significance tests.","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the empirical claims hinge on a single public dataset, and the loss-selection procedure is ambiguous. If the loss ablation used the test set, the SOTA claim would need to be re-evaluated on a properly held-out split. The authors are also authors of VQLIE [22], one of the compared baselines; this is not improper, but it should be stated explicitly for transparency. The manuscript fits the journal's scope, but the lack of code and the underspecified comparison protocol are significant reproducibility concerns that should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"BrightVAE is a plausible, incremental extension of hierarchical VQ-VAE for low-light endoscopic enhancement. The architecture is new in its combination: an attention-augmented encoder (Attencoder), attention-weighted quantization (Attenquant), dual-receptive-field processing, and skip connections. That is a legitimate extension of Peng et al. and the authors' own VQLIE, not a new paradigm, but it is a real attempt to adapt VQ-VAE to medical imaging.\n\nThe paper does useful empirical work: an ablation over eight similarity losses, a component study, and comparison against 14 existing methods on the public Endo4IE benchmark. The reported gains of about 3.25 dB PSNR over Mou et al. and 3.34 dB over plain VQ-VAE are substantial if true.\n\nThe soft spots are in the evidence, not the idea. First, the loss-function ablation in §4.D.2 selects the SSIM loss by looking at PSNR and SSIM scores, and the paper never states whether this was done on a held-out validation split. Section 4.B describes only train (690) and test (266) sets. If the test set was used to pick the loss, the Table 1 numbers are not an independent evaluation. That is a genuine, load-bearing concern. Second, no code, no error bars, no multiple seeds; the tables appear to be images, so numbers are not machine-readable and the protocol for re-running 14 baselines is unspecified. Third, the evaluation rests on one synthetic dataset. None of these make the method wrong; they make the SOTA claim unverified.\n\nI credit the authors for the breadth of the ablation and for situating the work in the literature. Self-citation is not an issue here; comparing to your own earlier VQLIE is normal.\n\nBottom line: for people working in low-light medical image enhancement, this is a relevant read and a reasonable baseline to test. It does not deserve a desk reject; the architecture and problem are legitimate. But a serious referee should ask for code and a clear validation protocol, and require the loss to be chosen on a validation split before test-set numbers are reported. I would send it to review with a request for major revision.","headline":"A credible incremental VQ-VAE extension with a strong-looking but under-validated SOTA claim; the loss ablation appears to use the test set, and the paper needs code and a validation split before the numbers can be trusted.","tokens_in":11480,"tokens_out":2658,"would_cite":false,"duration_ms":26526,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"BrightVAE, an attention-augmented hierarchical VQ-VAE, reports top PSNR, SSIM, and LPIPS for enhancing low-light endoscopic images on Endo4IE.","keywords":["low-light image enhancement","luminosity enhancement","endoscopic imaging","hierarchical VQ-VAE","attention mechanism","SSIM","LPIPS","Endo4IE dataset"],"falsifier":"Re-run the eight-way loss ablation from Section 4.D.2 with a held-out validation split, freeze the winning loss, retrain BrightVAE, and evaluate on the withheld 266-image Endo4IE test set; if a non-SSIM loss wins on validation or BrightVAE no longer leads every baseline, the central claim is not supported.","tokens_in":10510,"feed_emoji":"🩺","tokens_out":9626,"duration_ms":85112,"temperature":0.7,"pith_summary":"BrightVAE is an architecture for brightening underexposed endoscopic images, built on a hierarchical vector-quantized variational autoencoder (VQ-VAE). The paper claims that adding attention to the encoder and to the vector-quantization step, processing features through global and local receptive fields, and joining encoder to decoder with skip connections lets the model beat all compared methods on the Endo4IE benchmark. The reported margins are about 3.25 dB in PSNR and 0.045 in SSIM over the closest medical-imaging baseline, with a lower (better) LPIPS score as well. If these numbers hold, BrightVAE would give clinicians a clearer view of dark endoscopic scenes, and that visibility matters because diagnosis and treatment planning depend on what these images reveal.","feed_headline":"BrightVAE outshines 14 methods on dark endoscopy images","feed_subtitle":"Attention-based VQ-VAE lifts PSNR by 3.25 dB over prior best on Endo4IE.","key_machinery":"The load-bearing mechanism is the pair of attention-augmented modules inside a hierarchical VQ-VAE: the Attencoder and the Attenquant. The Attencoder appends a multi-head attention layer to a convolutional encoder so that the model can emphasize the image regions that matter most for brightness and detail. The Attenquant computes an attention score for each feature, normalizes the scores with softmax, and uses them to weight the squared Euclidean distances between encoder outputs and codebook embeddings before selecting the closest embedding, which biases quantization toward perceptually important features. These modules run in a dual-receptive-field design with global and local branches, and skip connections carry encoder feature maps directly to the decoder to preserve spatial detail. The total loss combines an MSE restoration term, a VQ latent-loss term with a stop-gradient operator, and an SSIM-based similarity term weighted 0.08.","core_discovery":"The central claim is that BrightVAE surpasses all compared models in PSNR, SSIM, and LPIPS for low-light endoscopic image enhancement on the Endo4IE dataset. This is achieved by a hierarchical VQ-VAE whose encoder (Attencoder) and quantizer (Attenquant) are augmented with attention, and whose global and local receptive-field branches are coupled through skip connections to the decoder. The paper reports a PSNR gain of about 3.252 dB, an SSIM gain of 0.045, and an LPIPS reduction of about 0.014 over the nearest medical-imaging competitor, and gains of 3.337 dB PSNR, 0.03 SSIM, and 0.038 LPIPS over the plain hierarchical VQ-VAE that inspired it. Those numbers are presented as evidence that the attention modules and dual-receptive design preserve fine details while correcting uneven illumination.","pith_inferences":["Because Endo4IE's low-light images are synthetic pairs produced by a pretrained generative model, the reported advantage should be rechecked on real endoscopic video before relying on it clinically.","The paper gives no inference-time cost or parameter count, so a deployment study measuring latency and memory on endoscopy hardware is a natural next step.","Since the SSIM loss won the ablation while a perceptual loss achieved the best secondary LPIPS value, combining SSIM with a perceptual term is a plausible variant the authors did not test."],"forward_implications":["On the Endo4IE benchmark, BrightVAE currently stands as the top-performing method across PSNR, SSIM, and LPIPS, assuming the reported test numbers are accurate.","The 3.34 dB gain over a plain hierarchical VQ-VAE indicates that the attention-augmented encoder and quantizer, not the backbone alone, produce most of the improvement.","The identified loss configuration, with reconstruction, latent, and SSIM similarity terms weighted 1.0, 0.25, and 0.08, gives a starting recipe for future endoscopic low-light enhancement methods.","After training, the method takes a single underexposed frame and produces an enhanced image without needing paired reference data at inference, so it is directly applicable to endoscopic video frames."],"supporting_citations":[{"why":"It supplies the Endo4IE dataset of 690 training and 266 test low-light and normal-light endoscopic image pairs used in every comparison.","marker":"[3]"},{"why":"It provides the hierarchical VQ-VAE backbone, the ConvResBlock, and the quantization routine that BrightVAE modifies with attention.","marker":"[23]"},{"why":"It defines the closest medical-imaging baseline against which BrightVAE's claimed 3.252 dB PSNR margin is measured.","marker":"[5]"},{"why":"It defines the SSIM measure that BrightVAE uses both as an evaluation metric and as the selected similarity loss in the ablation.","marker":"[24]"},{"why":"It defines the LPIPS perceptual metric used for evaluation and supplies the perceptual-loss candidate tested in the ablation.","marker":"[25]"},{"why":"It presents the earlier VQ-VAE-based low-light enhancement approach that motivates the VQ-VAE direction and is included as a comparative baseline.","marker":"[22]"}],"fun_headline_variants":["BrightVAE boosts dark endoscopy images by 3.25 dB PSNR","Attention-based VQ-VAE enhances low-light endoscopy images","Hierarchical VQ-VAE with attention improves endoscopy luminosity","BrightVAE lifts PSNR by 3.25 dB on dark endoscopy","BrightVAE: 3.25 dB PSNR gain for dark endoscopy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results stand on the assumption that the SSIM similarity loss was chosen without looking at the Endo4IE test set; if the test set guided that choice, the reported ranking is not an independent comparison.","fun_headline_variants_meta":{"raw":{"variants":["BrightVAE boosts dark endoscopy images by 3.25 dB PSNR","Attention-based VQ-VAE enhances low-light endoscopy images","Hierarchical VQ-VAE with attention improves endoscopy luminosity","BrightVAE lifts PSNR by 3.25 dB on dark endoscopy","BrightVAE: 3.25 dB PSNR gain for dark endoscopy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001053,"raw_usage":{"total_tokens":4440,"prompt_tokens":981,"completion_tokens":3459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":3358}},"tokens_in":597,"tokens_out":3459,"duration_ms":24021,"temperature":1.0,"reasoning_tokens":3358,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:02:52.346904+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the eight-way loss ablation from Section 4.D.2 with a held-out validation split, freeze the winning loss, retrain BrightVAE, and evaluate on the withheld 266-image Endo4IE test set; if a non-SSIM loss wins on validation or BrightVAE no longer leads every baseline, the central claim is not supported.","supporting_citations":[{"cited_title":"General”, and M refers to “Medical","cited_arxiv_id":null,"evidence_quote":"It supplies the Endo4IE dataset of 690 training and 266 test low-light and normal-light endoscopic image pairs used in every comparison."},{"cited_title":"Revealing Shadows: Low-Light Image Enhancement Using Self- Calibrated Illumination,","cited_arxiv_id":null,"evidence_quote":"It provides the hierarchical VQ-VAE backbone, the ConvResBlock, and the quantization routine that BrightVAE modifies with attention."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the closest medical-imaging baseline against which BrightVAE's claimed 3.252 dB PSNR margin is measured."},{"cited_title":"Beyond brightening low-light images,","cited_arxiv_id":null,"evidence_quote":"It defines the SSIM measure that BrightVAE uses both as an evaluation metric and as the selected similarity loss in the ablation."},{"cited_title":"EFINet: Restoration for low-light images via enhancement-fusion iterative network,","cited_arxiv_id":null,"evidence_quote":"It defines the LPIPS perceptual metric used for evaluation and supplies the perceptual-loss candidate tested in the ablation."},{"cited_title":"These diverse approaches underline a dynamic evolution of learning-based methods in image processing","cited_arxiv_id":null,"evidence_quote":"It presents the earlier VQ-VAE-based low-light enhancement approach that motivates the VQ-VAE direction and is included as a comparative baseline."}],"review_version":1}