{"id":"6ad6e2b3-5ea9-4d02-a6ce-062c7ad82b44","arxiv_id":"2508.04734","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Multi-level VQGAN does not consistently beat single-level VQGAN across two datasets for mIF-to-H&E virtual staining, despite the paper's claim of superiority.","lead":"A two-level VQGAN trained on paired multiplex immunofluorescence and H&E images produced virtual H&E stains that beat a conditional GAN on reconstruction and some downstream tasks on the Orion colorectal dataset, but the single-level VQGAN won on the CODEX dataset. The paper's headline claim that the multi-level VQGAN is a robust, superior architecture is not supported by its own results.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-hoc evaluation choices (CONCH used for both data curation and label metric; watershed selected after deep-learning segmenters fail) undermine the claim that the multi-level VQGAN is superior.","rationale":"I partially agree with the reader. The reader's weakest_assumption correctly identifies the CONCH double-use as a potential circularity, and this is part of the problem. However, I think the more general and more damaging issue is the post-hoc selection of the entire evaluation protocol: CONCH for both curation and semantic scoring, and watershed chosen only after deep-learning segmenters underperformed. These choices are explicitly documented in the paper (Sec. III.A, IV.B.1–2), so the concern is not speculative. The paper's own data do show that the 2-level VQGAN beats the cGAN on both functional tasks on both datasets, so the narrow claim 'VQGAN beats cGAN' is not contradicted. But the broader claim of 'robust and superior architecture' is not established because the metrics used to demonstrate superiority were not fixed in advance and may encode the authors' preferences. The absence of error bars or significance tests compounds this: the differences are often small (e.g., Orion L1 0.1491 vs 0.1545; CODEX IoU 0.8028 vs 0.8192), and any one of them could easily be noise. A simple re-analysis with a fixed protocol would determine whether the VQGAN advantage is real. Given that the paper presents this as a rigorous head-to-head comparison and the evaluation is not robust, rejection (or at best major revision with re-analysis) is warranted. I credit the paper for using two public paired datasets with patient-level splits and for comparing multiple architectures, but these strengths do not overcome the evaluation concerns.","tokens_in":9201,"tokens_out":8142,"duration_ms":95311,"concrete_test":"Fix the evaluation protocol before running it: (a) train/split on Orion without CONCH-based class balancing (or use a non-CONCH balancing method); (b) evaluate label consistency with a foundation model not used in curation (e.g., UNI or Virchow) instead of CONCH; (c) use a state-of-the-art deep-learning segmenter (e.g., Cellpose) as the primary segmentation metric, reporting watershed only as secondary; (d) run 3–5 random seeds and report mean±SD with paired significance tests. If the multi-level VQGAN no longer significantly outperforms the cGAN on these fixed metrics, the original claim is an artifact of evaluation choices.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim (Abstract; Conclusion) that a multi-level VQGAN is a 'robust and superior architecture' for mIF-to-H&E translation rests on two downstream evaluations. Both are vulnerable to post-hoc selection. (1) Semantic preservation (Sec. IV.B.1) is measured by label consistency using CONCH, but CONCH was also used in Sec. III.A to zero-shot classify and class-balance the Orion dataset. The test set is therefore enriched for patches CONCH labels confidently, and the metric rewards models whose outputs align with CONCH's specific decision boundary. The paper never examines this circularity. (2) Nuclei segmentation (Sec. IV.B.2) initially used Cellpose and StarDist, which 'showed poor agreement ... across all generative models'; the authors then state 'To enable a comparison, we instead used a traditional watershed algorithm.' This is an explicit switch of the evaluation metric after observing the first-choice metrics did not favor the models. No pre-registration or multiple-testing correction is reported. Additionally, no error bars, confidence intervals, or significance tests are provided for any table (Tables II–V), so apparent differences (e.g., Orion IoU 0.6800 vs 0.6376) may be within noise. Because the conclusion of superiority depends entirely on these confounded and post-hoc metrics, the central claim is not currently supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a paired, supervised approach to virtual H&E staining from multiplex immunofluorescence (mIF) images using a multi-level VQGAN, and compares it against a single-level VQGAN and a pix2pix-style cGAN on two public colorectal cancer datasets (Orion, CODEX). The authors report reconstruction metrics (L1, L2, SSIM, PSNR) and two downstream evaluations: tissue label consistency using a pretrained CONCH model, and nuclei segmentation using a watershed algorithm. The central claim is that the multi-level VQGAN is a 'robust and superior architecture' for generating scientifically useful virtual stains.","tokens_in":9540,"tokens_out":2974,"duration_ms":35734,"significance":"If the claimed superiority held, the paper would provide a useful benchmark for discrete-latent generative models in virtual staining, a topic with limited prior work for high-dimensional mIF-to-H&E translation. The study uses paired data with patient-level splits, which is a sound design for avoiding data leakage, and it contributes a systematic comparison of cGAN, VQGAN, and hierarchical VQGAN on two datasets. However, the evaluation has serious methodological flaws, including post-hoc selection of the segmentation metric, double use of the CONCH model for dataset curation and outcome measurement, and absence of any statistical uncertainty estimates. Because the central claim depends on these confounded and internally inconsistent results, the paper's significance is not currently established.","major_comments":[{"comment":"The abstract and conclusion claim that the multi-level VQGAN is a 'robust and superior architecture.' This is contradicted by the paper's own CODEX results: the single-level VQGAN achieves higher label consistency (53.5% vs 48.4%, Table IV) and higher watershed IoU (0.8192 vs 0.8028, Table V) than the multi-level VQGAN. The paper does not address this inconsistency, so the central superiority claim is unsupported.","section":"Abstract; Section IV.B; Tables IV and V"},{"comment":"The semantic-preservation metric uses CONCH to classify generated images, but CONCH was also used to curate the Orion training set via zero-shot classification and class balancing. This creates a circular dependency: the test set is enriched for patches CONCH classifies confidently, and the metric rewards outputs aligned with CONCH's specific decision boundary. The paper neither controls for this nor discusses its potential bias, making the Orion label-consistency advantage (69.9% vs 62.9%) unreliable as evidence of general semantic preservation.","section":"Section III.A and Section IV.B.1"},{"comment":"The authors state that Cellpose and StarDist 'showed poor agreement' and then switched to a traditional watershed algorithm 'to enable a comparison.' This is an explicit post-hoc change of the evaluation metric after observing the first-choice deep-learning segmenters did not yield usable results. No pre-registration, multiple-testing correction, or acknowledgment of this selection bias is provided. The motivating application is the H&E-based CAD ecosystem, which largely relies on deep-learning segmentation; watershed results do not demonstrate utility for that ecosystem.","section":"Section IV.B.2"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any metric. Differences such as Orion L1 (0.1491 vs 0.1545) and Orion IoU (0.6800 vs 0.6376) may be within noise, especially given patch-level evaluation and an unreported number of patients. Without variance estimates, the claimed improvements are not supported. This problem affects the reconstruction metrics as well as the downstream metrics.","section":"Tables II-V"}],"minor_comments":[{"comment":"The sentence 'Hematoxylin and Eosin (H&E) staining has served as the foundation... (H&E) [2]' has a duplicate '(H&E)' and an awkward break. Please revise.","section":"Introduction"},{"comment":"The terms '1-level' and 'single-level' are used interchangeably. Standardize terminology for consistency.","section":"Throughout"},{"comment":"The text says 'Our proposed 2-level VQGAN' while the abstract and conclusion use 'multi-level VQGAN.' Clarify whether the contribution is specifically the 2-level variant or the hierarchical approach generally.","section":"Section IV.A"},{"comment":"Figure 4 is referenced as showing box plots for Orion, but the figure is not present in the provided text. Ensure all figures are included and captions match the discussion.","section":"Section IV.B.2"},{"comment":"No information is given about code availability, hyperparameter search ranges, or the number of patients in the splits. Add these details to improve reproducibility.","section":"General"}],"recommendation":"reject","confidential_remarks":"The paper has merit as a benchmark study, but the central claim of multi-level VQGAN superiority is directly contradicted by the CODEX results, and the downstream evaluation is compromised by post-hoc metric selection and the CONCH double use. These issues are load-bearing and cannot be fully resolved without substantial reanalysis and reframing of the claims. The authors may be able to resubmit a revised manuscript with corrected claims and proper statistical treatment, but the current version does not meet the bar for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first paired mIF-to-H&E translation comparison I know of, and that alone makes it worth a look for anyone in virtual staining. But the \"multi-level VQGAN is superior\" claim in the abstract and conclusion does not survive the paper's own numbers. On CODEX, the single-level VQGAN beats the two-level model on both downstream tasks (53.5 vs 48.4 label consistency; 0.8192 vs 0.8028 nuclei IoU). The two-level model wins on Orion, so the honest conclusion is dataset-dependent, not architectural superiority.\n\nWhat the paper does well: it uses two public paired datasets with patient-level splits, tests downstream utility rather than only pixel metrics, and is transparent about the segmentation problem. The reconstruction tables mostly favor the two-level VQGAN, and the hierarchical quantization is a reasonable extension of VQGAN. The text also stays appropriately cautious about clinical readiness.\n\nThe soft spots are real. The label-consistency metric uses CONCH, and CONCH was also used to class-balance the Orion training set. That does not directly invalidate the model comparison, but it means the Orion semantic-preservation result partly rewards outputs that agree with CONCH's prior. The bigger problem is the post-hoc switch to watershed after Cellpose and StarDist \"showed poor agreement across all generative models.\" Selecting the metric after seeing the first-choice metrics fail is the kind of multiple-comparison artifact that referee reports exist for. And there are no confidence intervals or significance tests anywhere, so differences like 0.6800 vs 0.6270 on Orion may be noise.\n\nNone of this means the paper is worthless. It is a legitimate new application and a reasonable baseline for future work. But it needs major revision before it can claim what it claims. I'd send it to peer review, not desk-reject it, with the expectation that the authors either soften the superiority claim to match the CODEX results or provide much stronger evidence. I'd also ask for error bars and a pre-specified segmentation evaluation. As written, the central claim is overreach; the empirical comparison itself is worth keeping.","headline":"Useful first paired mIF-to-H&E comparison, but the 'superior architecture' claim is not supported by the paper's own CODEX results and the post-hoc evaluation choices are too convenient.","tokens_in":10013,"tokens_out":2950,"would_cite":false,"duration_ms":39606,"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":"A multi-level VQGAN trained on paired multiplex immunofluorescence and H&E images generates virtual H&E stains that outperform a standard conditional GAN on downstream nuclei segmentation and tissue-label consistency, on two colorectal canc","keywords":["virtual staining","multiplex immunofluorescence","VQGAN","vector quantization","conditional GAN","computational pathology","nuclei segmentation","colorectal cancer"],"falsifier":"Retrain the cGAN and VQGAN on the Orion data with random class-balanced sampling instead of CONCH-based curation, and score semantic preservation with an independent tissue classifier or pathologist panel. If the VQGAN's label-consistency and segmentation margins persist, the architecture claim holds; if they shrink or reverse, the reported advantage is an artifact of the shared CONCH evaluator. A complementary check is sweeping the number of quantization levels on the 58-channel CODEX data to see whether the optimal hierarchy depth depends on input dimensionality.","tokens_in":9096,"feed_emoji":"🔬","tokens_out":12375,"duration_ms":124122,"temperature":0.7,"pith_summary":"This paper sets out to establish that a multi-level VQGAN—a generative model whose latent space is a hierarchy of discrete codebooks—produces virtual H&E stains from multiplex immunofluorescence (mIF) images that are more useful for automated pathology analysis than the stains produced by a standard continuous-latent conditional GAN. The authors test this on paired mIF/H&E patches from two colorectal cancer datasets, judging usefulness not by visual appeal alone but by downstream nuclei segmentation overlap and by whether a pre-trained tissue labeler assigns the same labels to generated and ground-truth H&E. On both datasets, at least one VQGAN variant beats the cGAN on every reported downstream metric, and the two-level VQGAN leads on the 19-channel Orion dataset. If the finding is right, the discrete-latent architecture is a viable route for applying H&E-based computer-aided diagnosis tools to high-dimensional mIF data without physical restaining.","feed_headline":"VQGAN virtual stains beat cGAN for downstream pathology tasks","feed_subtitle":"Discrete hierarchical codes preserve tissue; nuclei segmentation and label consistency improve on two CRC datasets.","key_machinery":"The central object is the multi-level VQGAN: a convolutional encoder maps an mIF patch into feature maps at two spatial resolutions, each quantized by its own learned codebook of 1024 entries, and a decoder reconstructs the H&E image from the discrete codes. The hierarchy is intended to let the low-resolution codebook carry coarse tissue architecture while the high-resolution codebook carries cellular detail. Training uses a composite loss $L_{\\mathrm{total}} = \\lambda_{\\mathrm{recon}} L_{\\mathrm{recon}} + \\lambda_{\\mathrm{perc}} L_{\\mathrm{perc}} + \\lambda_{\\mathrm{style}} L_{\\mathrm{style}} + \\lambda_{\\mathrm{stain}} L_{\\mathrm{stain}} + \\lambda_{\\mathrm{adv}} L_{\\mathrm{adv}}$, where the","core_discovery":"The paper's central claim is that hierarchical vector quantization can compress a 19- or 58-channel mIF input into discrete codes that decode into an H&E image whose scientific content matches ground-truth H&E better than a standard conditional GAN with a continuous latent space. The decisive evidence is functional, not perceptual: using a watershed segmenter, the two-level VQGAN reaches a mean IoU of 0.6800 on Orion versus 0.6376 for the cGAN; using CONCH as a tissue labeler, generated images agree with ground-truth labels 69.9% of the time for the two-level VQGAN versus 62.9% for the cGAN. On the 58-channel CODEX dataset, the single-level VQGAN leads with 53.5% label consistency and 0.8192","pith_inferences":["A consequence the authors leave implicit: the CONCH model is used twice, to balance the Orion training set and to score label consistency, so a direct test with independently curated patches and an independent semantic labeler is needed to know whether the VQGAN's margin is about tissue fidelity or alignment with CONCH's preferences.","On the 58-channel CODEX data, the single-level VQGAN beats the two-level model on both label consistency and segmentation, suggesting the optimal number of quantization levels is dataset-dependent; a systematic sweep of codebook levels against channel depth would test this.","The failure of Cellpose and StarDist to transfer between ground-truth and generated H&E suggests current virtual stains preserve low-frequency layout but not all high-frequency textural cues deep segmenters rely on; improving those cues is a concrete next target.","The winning label-consistency scores (roughly 50-70%) mean virtual H&E is not yet interchangeable with physical H&E for diagnosis; the paper's own caution about clinical readiness leaves open how much agreement a diagnostic workflow would require."],"forward_implications":["If the claim is correct, mIF images can be rendered as H&E-like views quickly enough to give morphological context without waiting for physical histology, and existing H&E-based CAD pipelines become applicable to mIF data.","Discrete latent codes are sufficiently expressive that a finite codebook does not hurt a high-dimensional (19-58 channel) image translation task.","For class-balanced, 19-channel data, adding a second quantization level improves reconstruction and downstream utility over a single level.","VQGAN-generated virtual stains support watershed nuclei segmentation well enough that a traditional segmenter transfers better than deep segmentation models.","The learned discrete codes are proposed as a possible foundation for other mIF downstream tasks, such as cell classification or tumor-microenvironment clustering."],"supporting_citations":[{"why":"Supplies the Orion dataset of paired 19-channel mIF and H&E images used for the main comparison.","marker":"[1]"},{"why":"Supplies the CODEX dataset of paired 58-channel mIF and H&E images used to test generalization to deeper mIF.","marker":"[19]"},{"why":"Provides CONCH, the foundation model used both to balance the Orion training set and to measure tissue-label consistency.","marker":"[20]"},{"why":"Introduces the VQGAN architecture and codebook training that the paper's single- and multi-level models build on.","marker":"[11]"},{"why":"Defines the conditional GAN architecture used as the supervised baseline.","marker":"[16]"},{"why":"Motivates the multi-level latent hierarchy with separate codebooks at different spatial resolutions.","marker":"[12]"},{"why":"Introduces discrete latent variable models that ground the vector-quantization approach.","marker":"[10]"}],"fun_headline_variants":["VQGAN virtual stains outdo cGAN for pathology tools","Hierarchical VQGAN beats cGAN on virtual H&E utility","Virtual H&E from mIF: VQGAN wins on functional tasks","VQGAN virtual stains: better for nuclei segmentation","Discrete codes yield more useful virtual H&E stains"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that using the same CONCH model to balance the Orion training patches and to score label consistency does not bias the architecture comparison; if CONCH's preferences shaped which patches were learned, the label-consistency metric may reward VQGAN outputs for matching CONCH's biases rather than true tissue semantics.","fun_headline_variants_meta":{"raw":{"variants":["VQGAN virtual stains outdo cGAN for pathology tools","Hierarchical VQGAN beats cGAN on virtual H&E utility","Virtual H&E from mIF: VQGAN wins on functional tasks","VQGAN virtual stains: better for nuclei segmentation","Discrete codes yield more useful virtual H&E stains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1119,"prompt_tokens":831,"completion_tokens":288,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":203}},"tokens_in":575,"tokens_out":288,"duration_ms":3462,"temperature":1.0,"reasoning_tokens":203,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T01:01:42.737262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the cGAN and VQGAN on the Orion data with random class-balanced sampling instead of CONCH-based curation, and score semantic preservation with an independent tissue classifier or pathologist panel. If the VQGAN's label-consistency and segmentation margins persist, the architecture claim holds; if they shrink or reverse, the reported advantage is an artifact of the shared CONCH evaluator. A complementary check is sweeping the number of quantization levels on the 58-channel CODEX data to see whether the optimal hierarchy depth depends on input dimensionality.","supporting_citations":[{"cited_title":"High-plex immunofluorescence imaging and traditional histology of the same tissue section for discovering image-based biomarkers,","cited_arxiv_id":null,"evidence_quote":"Supplies the Orion dataset of paired 19-channel mIF and H&E images used for the main comparison."},{"cited_title":"Coordinated cellular neighborhoods orchestrate antitumoral immunity at the colorectal cancer invasive front,","cited_arxiv_id":null,"evidence_quote":"Supplies the CODEX dataset of paired 58-channel mIF and H&E images used to test generalization to deeper mIF."},{"cited_title":"Taming transformers for high- resolution image synthesis,","cited_arxiv_id":null,"evidence_quote":"Introduces the VQGAN architecture and codebook training that the paper's single- and multi-level models build on."},{"cited_title":"Generating diverse high-fidelity images with vq-vae-2,","cited_arxiv_id":null,"evidence_quote":"Motivates the multi-level latent hierarchy with separate codebooks at different spatial resolutions."},{"cited_title":"Neural discrete representation learning,","cited_arxiv_id":null,"evidence_quote":"Introduces discrete latent variable models that ground the vector-quantization approach."}],"review_version":1}