{"id":"1f0592d2-197a-433e-af8c-7184cf8f4341","arxiv_id":"2506.07559","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CCPL aligns nucleus and membrane channel statistics using Gigapath features and FOD maps, reporting improved metrics on BCI and MIST-HER2.","lead":"Virtual staining converts H&E stained tissue into HER2-stained images without extra chemicals. This paper adds losses that align cell nucleus and membrane features across channels, reporting quality improvements on two public datasets, though one FID claim contradicts its own table.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Table I contradicts the stated SOTA FID and PCC claims on BCI, so the central quantitative claim is internally inconsistent.","rationale":"The reader's weakest assumption concerned the biological validity of the NMCC correlation, but the more immediate and checkable problem is that the paper's own Table I contradicts the SOTA claim. A claim based on quantitative superiority cannot stand when the stated numbers show two baselines with lower FID and one with higher PCC on BCI. The FID issue is arithmetic, not a question of interpretation. The human-evaluation evidence is also weak because discordant scores were excluded and the cited supporting reference [17] is an unrelated social-dialogue paper, but the table contradiction alone is load-bearing. The proposed method may still be useful, and the PSNR/SSIM improvements are real, so the paper might be acceptable after correction; this does not change the reader's conditional verdict, though the required revisions should include fixing the FID/PCC claims and either removing or properly supporting reference [17].","tokens_in":8910,"tokens_out":8642,"duration_ms":98512,"concrete_test":"Independently recompute FID and PCC on the BCI test set using the authors' released or retrained model and the same preprocessing as Table I, comparing against ASP and PSPStain. If the recomputed values match Table I, then Section IV-C's FID statement is false and the SOTA claim must be narrowed; if they differ, the reported table is unreliable. This directly settles whether the central quantitative claim is valid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section IV-C, the text states 'CCPL achieves the lowermost FID scores on the BCI dataset,' but Table I reports FID=57.81 for Ours versus 54.72 for ASP and 48.42 for PSPStain; since lower FID is better, CCPL is not lowest. The abstract also claims SOTA on PCC, but Table I shows ASP has PCC=0.2609 vs Ours=0.2441 on BCI. The central claim that CCPL 'achieves state-of-the-art performance on quantitative metrics and human evaluations on the BCI and MIST-HER2 datasets' is therefore not supported by the paper's own results. The human evaluation also excludes discordant scores and cites an unrelated reference [17], but the quantitative contradiction is sufficient to undermine the strongest claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Cross-Channel Perception Learning (CCPL) for H&E-to-IHC virtual staining. The method decomposes HER2 immunohistochemistry into hematoxylin (nucleus) and DAB (membrane) channels, and introduces three auxiliary losses: a Dual-Channel Perception (DCP) loss based on Focal Optical Density statistics, a Feature Distillation (FD) loss that aligns generated and real image features extracted by Gigapath's Tile Encoder, and a Nucleus-Membrane Cross-Channel Correlation (NMCC) loss. Experiments are reported on the BCI and MIST-HER2 datasets using PSNR, SSIM, PCC, FID, and pathologist ratings, with the paper claiming state-of-the-art performance.","tokens_in":9033,"tokens_out":2805,"duration_ms":35806,"significance":"If the claims were fully supported, the idea of exploiting cross-channel nucleus-membrane correlations with a pathology foundation model would be a useful contribution to virtual staining, and the lightweight distillation approach is practically attractive. The paper is commendable for giving explicit loss definitions, using ground-truth-supervised objectives, and including an ablation study with human evaluation. However, the central quantitative claim is contradicted by the paper's own Table I, and the NMCC loss as written is technically questionable, so the significance as currently stated cannot be accepted without revision.","major_comments":[{"comment":"The text states that \"CCPL achieves the lowermost FID scores on the BCI dataset,\" but Table I reports FID 57.81 for Ours versus 48.42 for PSPStain and 54.72 for ASP; since lower FID is better, this statement is false. The abstract and contribution list also claim state-of-the-art PCC, but Table I shows ASP has PCC 0.2609 versus 0.2441 for Ours on BCI. The central claim that CCPL achieves state-of-the-art quantitative performance is therefore not supported by the paper's own results, and the affected statements must be corrected and the claims appropriately qualified.","section":"Section IV-C, Table I"},{"comment":"R_g and R_r are defined as scalars in Eqs. (7) and (8): each is a weighted sum of an L2 norm and a cosine distance, both of which are scalar quantities. Eq. (9) then computes an L2 norm and a cosine similarity between R_g and R_r. A cosine similarity between scalars is degenerate (it can only be +1 or -1 depending on sign), so the second term of L_cross is either zero or a constant, and the loss reduces to a scalar L2 distance. This makes the proposed cross-channel consistency loss poorly motivated and its gradient behavior unclear. The authors should either define R as a vector-valued correlation representation or justify why a scalar distance is sufficient, and provide an analysis of the loss landscape.","section":"Section III-C, Eqs. (7)-(9)"},{"comment":"The human evaluation protocol is not reproducible as described. The text states that \"images with significant scoring discrepancies were excluded\" but does not specify the discrepancy threshold, the number of excluded images, the number of images scored, or how the three pathologists' scores were aggregated. In addition, the human evaluation is cited to reference [17], which is an arXiv paper on social conversational dialogues and appears unrelated to pathology or image evaluation. Since human evaluation is part of the paper's central claim, full details and an appropriate citation are required.","section":"Section IV-A, Human Evaluation"},{"comment":"The NMCC premise that the feature-space distance and cosine similarity between Gigapath features of the hematoxylin and DAB channels capture the biologically meaningful nucleus-membrane relationship is not validated. The paper does not show that minimizing the difference between R_generated and R_real actually enforces spatial or intensity correlations that are pathologically meaningful; it could simply be aligning arbitrary feature statistics. A concrete validation, such as correlation with HER2 scoring annotations or a comparison against a simpler cross-channel pixel-space alignment, is needed to support the central contribution.","section":"Section III-C"}],"minor_comments":[{"comment":"The notation in Eq. (1) is confusing: \"F ODC = ((ODC)αC, if F O D C > T_C, 0 otherwise\" appears to use F ODC both as the output and in the condition; it should presumably be ODC > T_C. Please clarify.","section":"Eq. (1)"},{"comment":"There are several typographical errors, including \"demote\" for \"denote\" (Eq. (2)), \"groud truth\" (Fig. 2 caption), \"evlautions\" and \"bechmark\" in the abstract/conclusion areas, and \"Mode 3 vs. Model 4\" in Section IV-D that should be \"Model 3 vs. Model 4\".","section":"Throughout"},{"comment":"The ablation discussion attributes the FID decline for NMCC to the Inception feature extractor, but the full model still has a much higher FID (57.81) than the FD-only model (40.05), even though FD-only has the best FID overall. The text should more clearly acknowledge that the final model is not FID-optimal and explain this trade-off with respect to the SOTA FID claim in Section IV-C.","section":"Section IV-D, Table III"},{"comment":"The definitions of PCPA and SIA should state exactly which HER2 scoring guideline is used and how the \"proportion of positive cell nuclei\" is determined, since HER2 positivity is based on membrane staining rather than nuclear positivity in the standard clinical scoring.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea and the ablation study is informative, but the internally contradictory quantitative claims and the technically degenerate formulation of the NMCC loss require substantial revision. The authors should also replace the unrelated self-citation [17], which raises a credibility concern that the editor may wish to verify."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: the core idea is a new loss term for H&E-to-IHC virtual staining that measures cross-channel correlation between nucleus (H) and membrane (D) features, extracted by a pathology foundation model. That is a reasonable direction, and the ablations show each component contributes something. But the paper overstates its results, and the text contradicts its own table.\n\nThe new element is the NMCC loss, which aligns the feature-space distance between the H and D channels of the generated image to that of the real image. It is well-specified, supervised by ground truth, and does not add inference cost. The FOD-based dual-channel perception and Gigapath distillation are extensions of existing work, but the combination is new. The ablation study is genuinely useful: FD improves FID, NMCC improves SSIM/PCC, and the full model wins on several metrics on MIST-HER2 and on SSIM/PSNR on BCI.\n\nThe soft spots are real and not minor. Section IV-C says CCPL achieves the lowermost FID on BCI, but Table I shows Ours=57.81 while PSPStain=48.42 and ASP=54.72. It also says PCC ranks first and second on BCI and MIST-HER2 respectively, but on BCI ASP has 0.2609 vs Ours 0.2441—so the order is wrong. These are not typos in isolation; they directly support the abstract's SOTA claim. There are also no error bars or significance tests, the human evaluation protocol excludes discordant scores and cites an unrelated reference [17], and no code is released. The NMCC premise—that feature-space distance between H and D captures biological nucleus-membrane relationships—is plausible but not validated against any pathology annotation.\n\nNone of this kills the method. The losses are sensible, the ablations are informative, and on MIST-HER2 the model is genuinely competitive. But the current draft does not support the SOTA claim as written. The fix is straightforward: correct the text, add variance or significance, and clarify the human evaluation.\n\nThis paper is for people working on stain translation and virtual staining in computational pathology. It is a solid engineering contribution with a testable idea, not a conceptual breakthrough. It deserves peer review, but a strict referee should send it back for major revisions. I would not cite it in its current form.","headline":"Plausible loss combination and honest ablations, but the paper's own Table I contradicts the SOTA claim—must be corrected before the results can be trusted.","tokens_in":9623,"tokens_out":1993,"would_cite":false,"duration_ms":24943,"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":"Modeling the nucleus-membrane correlation during H&E-to-IHC translation yields virtual HER2 images that preserve pathologist-scored staining features better than earlier methods.","keywords":["virtual staining","H&E-to-IHC","HER2","cross-channel correlation","feature distillation","pathology foundation model","focal optical density","digital pathology"],"falsifier":"A concrete check: take a set of real HER2 images with pathologist-assigned membrane-intensity and positive-cell-proportion scores, compute $R$ for each, and see whether $R$ orders the images by score. If $R$ is uncorrelated with the scores, the NMCC target is not biologically meaningful. A complementary ablation would remove only the NMCC loss and have pathologists blind-compare the outputs; if the scores do not move, the cross-channel term is not the source of the reported fidelity gain.","tokens_in":8665,"feed_emoji":"🔬","tokens_out":10483,"duration_ms":118157,"temperature":0.7,"pith_summary":"This paper seeks to establish that the limited fidelity of H&E-to-IHC virtual staining comes in part from ignoring the relationship between cell nuclei and cell membranes, and that a loss function explicitly encoding that relationship fixes it. CCPL decomposes real HER2 images into hematoxylin (nucleus) and DAB (membrane) channels, uses a pathology foundation model's tile encoder to compare generated and real channels, and adds three losses: dual-channel perception on focal optical density statistics, feature distillation, and nucleus-membrane cross-channel correlation. On the BCI and MIST-HER2 benchmarks, CCPL reports higher pixel-level metrics and higher pathologist ratings for positive-cell proportion and staining-intensity accuracy than the compared methods. If the claim is right, virtual HER2 staining can reproduce the features pathologists actually score, which matters because chemical IHC staining is slow and costly and digital pathology would benefit from reliable surrogate stains.","feed_headline":"Nucleus-membrane pairing lifts virtual HER2 stain fidelity","feed_subtitle":"A cross-channel loss ties nuclei to membranes and beats prior H&E-to-IHC methods on two breast cancer benchmarks.","key_machinery":"The load-bearing object is the Nucleus-Membrane Cross-Channel Correlation (NMCC) loss. For a generated or real HER2 tile, CCPL first uses color deconvolution to isolate the hematoxylin and DAB channels and computes their separate Gigapath features; the correlation between the two channels is then summarized by a single scalar per image, $R = \\gamma \\|f_H - f_D\\|_2^2 + (1-\\gamma)(1-\\cos(f_H, f_D))$, combining an L2 distance and a cosine term. The NMCC loss compares the generated image's $R_g$ with the real image's $R_r$ in the same mixed L2-plus-cosine form. This scalar is what forces the generator to reproduce the spatial and intensity relationship between nuclei and membranes rather than matching each channel independently. The same Gigapath features also feed the feature-distillation loss, so one frozen encoder carries both the semantic-alignment and the cross-channel targets.","core_discovery":"The central claim, stated on the paper's own terms, is that cross-channel semantic consistency between the nucleus and the membrane is a learnable target for H&E-to-HER2 translation. CCPL obtains single-channel images by color deconvolution and focal optical density conversion, extracts features $f_{H,g}, f_{D,g}$ from the generated image and $f_{H,r}, f_{D,r}$ from the real image using Gigapath's tile encoder, and defines the cross-channel correlation $R = \\gamma \\|f_H - f_D\\|_2^2 + (1-\\gamma)(1 - \\cos(f_H, f_D))$ for each image. The nucleus-membrane cross-channel loss then enforces $R_g \\approx R_r$, alongside a feature-distillation loss that aligns generated and real representations in the same encoder space and a dual-channel perception loss that matches focal optical density statistics. The paper argues that this combination preserves the proportion of HER2-positive cells and the membrane staining intensity that HER2 scoring depends on, and reports quantitative and pathologist-evaluation evidence that it does so better than the compared prior methods.","pith_inferences":["Editorial inference: the same NMCC loss could be carried over to other membrane-localized IHC markers, such as EGFR or CD20, where positive membranes wrap around nuclei, because the loss itself does not depend on HER2-specific biology.","Editorial inference: the reported FID drop when NMCC is added suggests a metric-level ambiguity; a pathology-specific distribution metric would tell whether NMCC genuinely hurts distribution-level realism or only Inception-based FID.","Editorial inference: the color deconvolution thresholds and exponents ($T_H, T_D, \\alpha_H, \\alpha_D$) are fixed hyperparameters; testing CCPL across scanners and antibody kits would show whether the decomposition, and therefore the correlation target, stays stable."],"forward_implications":["If CCPL is correct, virtual HER2 images will preserve the two features pathologists score\\u2014membrane intensity and positive-cell proportion\\u2014more faithfully than earlier translation methods.","Because the extra losses use a frozen foundation-model encoder at training time only, the deployed generator remains lightweight and fast.","Each component has a distinct reported role: feature distillation improves distribution-level realism, NMCC improves structural correlation, and dual-channel perception improves pixel-level detail.","The method is demonstrated on two breast-cancer H&E-to-IHC benchmarks, so the gains are not tied to a single dataset."],"supporting_citations":[{"why":"Grounds the HER2 scoring requirement (membrane intensity and positive-cell proportion) that motivates the cross-channel design.","marker":"[12]"},{"why":"Supplies the Focal Optical Density formulation and the PSPStain baseline whose protein-aware strategy CCPL extends.","marker":"[14]"},{"why":"Provides the pretrained Gigapath Tile Encoder used for feature distillation and for measuring cross-channel correlation.","marker":"[16]"},{"why":"Provides the Pix2pix baseline in the H&E-to-IHC comparison and the stain-to-stain translation setting.","marker":"[25]"},{"why":"Introduces the BCI dataset and the Pyramidpix2pix baseline used for training and evaluation.","marker":"[26]"},{"why":"Introduces the ASP method and the MIST-HER2 subset that CCPL compares against.","marker":"[27]"},{"why":"Provides the CUT baseline and contrastive learning formulation on which the generator is built.","marker":"[29]"}],"fun_headline_variants":["Cross-channel nuclei-membrane learning boosts virtual HER2 staining","Teaching H&E-to-IHC models to link nuclei and membranes","HER2 virtual staining improved by cross-channel perception","Nucleus-membrane correlation loss sharpens virtual IHC","CCPL: Cross-channel learning for better H&E-to-HER2"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument has a single load-bearing premise: the feature-space difference between the hematoxylin and DAB channels, condensed into $R$, corresponds to the spatial and intensity relationship between nuclei and membranes that a pathologist reads; if that correspondence fails, the NMCC loss is aligning the generated and real images on a quantity unrelated to diagnosis.","fun_headline_variants_meta":{"raw":{"variants":["Cross-channel nuclei-membrane learning boosts virtual HER2 staining","Teaching H&E-to-IHC models to link nuclei and membranes","HER2 virtual staining improved by cross-channel perception","Nucleus-membrane correlation loss sharpens virtual IHC","CCPL: Cross-channel learning for better H&E-to-HER2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000522,"raw_usage":{"total_tokens":2560,"prompt_tokens":1013,"completion_tokens":1547,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":1463}},"tokens_in":629,"tokens_out":1547,"duration_ms":12227,"temperature":1.0,"reasoning_tokens":1463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:30:15.687292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: take a set of real HER2 images with pathologist-assigned membrane-intensity and positive-cell-proportion scores, compute $R$ for each, and see whether $R$ orders the images by score. If $R$ is uncorrelated with the scores, the NMCC target is not biologically meaningful. A complementary ablation would remove only the NMCC loss and have pathologists blind-compare the outputs; if the scores do not move, the cross-channel term is not the source of the reported fidelity gain.","supporting_citations":[{"cited_title":"Her2 status in breast cancer: changes in guidelines and complicating factors for interpretation,","cited_arxiv_id":null,"evidence_quote":"Grounds the HER2 scoring requirement (membrane intensity and positive-cell proportion) that motivates the cross-channel design."},{"cited_title":"Pathological semantics-preserving learning for h&e-to- ihc virtual staining,","cited_arxiv_id":null,"evidence_quote":"Supplies the Focal Optical Density formulation and the PSPStain baseline whose protein-aware strategy CCPL extends."},{"cited_title":"A whole-slide foundation model for digital pathology from real-world data,","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained Gigapath Tile Encoder used for feature distillation and for measuring cross-channel correlation."},{"cited_title":"Pix2pix-based stain-to-stain translation: A solution for robust stain normalization in histopathology images analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the Pix2pix baseline in the H&E-to-IHC comparison and the stain-to-stain translation setting."},{"cited_title":"Bci: Breast cancer immunohistochemical image generation through pyramid pix2pix,","cited_arxiv_id":null,"evidence_quote":"Introduces the BCI dataset and the Pyramidpix2pix baseline used for training and evaluation."},{"cited_title":"Adaptive supervised patchnce loss for learning h&e-to-ihc stain translation with inconsistent groundtruth image pairs,","cited_arxiv_id":null,"evidence_quote":"Introduces the ASP method and the MIST-HER2 subset that CCPL compares against."},{"cited_title":"Con- trastive learning for unpaired image-to-image translation,","cited_arxiv_id":null,"evidence_quote":"Provides the CUT baseline and contrastive learning formulation on which the generator is built."}],"review_version":1}