{"id":"db189fe0-adc6-4b8b-8391-f1dea4952f27","arxiv_id":"2505.20746","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Ui2i uses a CycleGAN-like model with U-Net generators, spectral normalization, and attention to preserve content, enabling cross-stain segmentation and unpaired separation of same-fluorophore biomarkers in microscopy images.","lead":"Researchers introduce Ui2i, an unpaired image-to-image translation model that transfers textures between medical image styles while preserving the shape and layout of biological structures. It enables nuclear segmentation across staining domains and, for the first time, separates two biomarkers that share the same fluorescent label using real, unpaired microscopy images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central novelty claim is not validated for the intended same-fluorophore regime: quantitative results come from a spectrally distinct surrogate (HT-T24, Sec. 8/Appendix G), while real same-fluorophore outputs are only qualitative; a same-fluorophore quantitative check is needed.","rationale":"The paper is a coherent engineering contribution, and I credit the authors for honestly documenting the HT-T24 mismatch in Appendix G and the hallucination failure in Appendix H. The reader's weakest assumption was the shared-content-space assumption and its unverified validity on real same-fluorophore data; I agree that this is important. My concern is adjacent but sharper: even if content spaces overlap, unpaired translation cannot uniquely fix a one-to-two-channel decomposition, because the cycle loss only requires that the two output channels reconstruct the input. The quantitative surrogate uses a different acquisition physics, and the real same-fluorophore results have no quantitative anchor. The DAPI-based external readout is feasible because DAPI is routinely acquired with these images and provides an independent nuclear mask. If the unmixed Ki67 channel does not match DAPI nuclei, then the model is not separating the markers on the actual intended data, and the 'first capable' claim should not stand. This does not change the verdict from the reader's conditional acceptance, but it tightens the condition: release code/data and supply a same-fluorophore quantitative check, not only the HT-T24 surrogate.","tokens_in":17514,"tokens_out":9426,"duration_ms":102133,"concrete_test":"On the real breast/colon same-fluorophore dataset, add a DAPI-based external quantitative readout: segment DAPI-positive nuclei, compute Dice/F1 between the DAPI mask and the Ki67 channel produced by Ui2i, and measure membrane localization of the E-cadherin channel (e.g., distance-to-nuclear-boundary profile or correlation with a membrane marker). Pre-specify a tolerance relative to the same readout on the HT-T24 surrogate outputs, for example requiring the Ki67-DAPI Dice to be within 0.1 of the surrogate value. If the real-data unmixing falls outside this tolerance, the central claim is unsupported for its intended regime; if it passes, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that Ui2i is the first to separate superimposed same-fluorophore IF signals from real unpaired data. That claim requires the learned one-to-two-channel decomposition to be correct on the intended acquisition regime, not merely cycle-consistent. The paper itself (Sec. 3) notes that infinitely many cycle-consistent generator pairs exist; for unmixing, any pair of output channels that reconstructs the input sum satisfies the cycle, so the adversarial and contrastive losses must supply the missing constraint. The only quantitative evidence (Sec. 8, Table 2) uses HT-T24, where SOX2 and Grasp65 are tagged with spectrally distinct fluorophores and the multiplexed image comes from sequential acquisition plus simultaneous excitation. Appendix G concedes these images 'do not faithfully reproduce the signal characteristics, noise patterns, or artifact profiles of true single-fluorophore multiplexing.' The real same-fluorophore breast/colon results (Figs. 5, 11, 12) are qualitative only; the text states ground truth is impossible. Appendix H shows that when source content is absent from the target domain, the model hallucinates structures. Therefore the observed HT-T24 performance may be an artifact of separating two distinct fluorophore statistics, and the same-fluorophore capability that defines the novelty is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Ui2i, an unpaired bidirectional image-to-image translation model built on CycleGAN, with U-Net generators with skip connections, removal of feature normalization in favor of approximate bidirectional spectral normalization, channel/spatial attention, scale augmentation, and a cross-domain contrastive loss. The model is evaluated on two biomedical tasks: (i) IHC-to-H&E translation followed by StarDist nuclear segmentation, where improvements over CycleGAN and over direct InstanSeg application are reported; and (ii) unmixing a single-channel immunofluorescence image into two channels representing two markers labeled with the same fluorophore, where quantitative results are reported on the HT-T24 dataset and qualitative results are shown for real breast/colon same-fluorophore data. The paper claims to be the first approach capable of separating superimposed IF signals using real, unpaired training data, and claims significantly greater content fidelity than existing i2i translation methods.","tokens_in":17729,"tokens_out":3996,"duration_ms":39988,"significance":"If the central claims are substantiated, the work would be practically significant: it would enable marker separation for same-fluorophore multiplexed immunofluorescence without paired training data, and it would provide a domain-adaptation pipeline for nuclear segmentation that outperforms direct application of a domain-pretrained segmenter. Strengths include evaluation with an external pretrained segmenter (StarDist), ablations of key architectural choices, and use of public datasets. However, the manuscript's load-bearing claim of same-fluorophore unmixing is not quantitatively validated in the intended regime; the only quantitative unmixing experiment uses spectrally distinct fluorophores that, as Appendix G concedes, do not reproduce same-fluorophore signal characteristics. The segmentation claim of 'significantly greater' fidelity is also not supported by statistical significance tests, given overlapping standard deviations in Table 1.","major_comments":[{"comment":"The quantitative evaluation of unmixing is performed on HT-T24, where SOX2 and Grasp65 are labeled with spectrally distinct fluorophores and the multiplexed image was acquired by simultaneous excitation; Appendix G explicitly states that these images 'do not faithfully reproduce the signal characteristics, noise patterns, or artifact profiles of true single-fluorophore multiplexing.' The central claim in the abstract and Section 8—that Ui2i is the first to separate superimposed signals in IF images using real, unpaired data—therefore remains unverified for the intended same-fluorophore regime. The real breast/colon results (Figs. 5, 11, 12) are qualitative only, with the text noting that ground truth is impossible. I request either a quantitative same-fluorophore evaluation (for example, synthetic mixtures created by adding real single-fluorophore channels from the same sample, with realistic noise and artifact injection) or a clear restriction of the novelty claim to the surrogate regime.","section":"Section 8, Table 2, Appendix G"},{"comment":"The paper correctly notes that infinitely many cycle-consistent generator pairs exist. For the unmixing task the degeneracy is even more direct: any pair of output channels whose sum equals the input satisfies the cycle consistency loss, so the adversarial and contrastive losses alone must select the intended biological decomposition. The manuscript does not provide a test isolating whether these losses break the degeneracy in the same-fluorophore regime; separating two spectrally distinct fluorophore statistics on HT-T24 could succeed without the model learning a same-fluorophore decomposition. A control experiment—for example, training on synthetic same-fluorophore mixtures with known ground truth, or checking that the two output channels sum to the input while matching per-channel distributions—is needed to make the unmixing claim load-bearing.","section":"Section 3, Eq. (3)"},{"comment":"The claim that Ui2i preserves image content fidelity 'to a significantly greater extent' than existing methods is not supported by any statistical significance test. In Table 1, standard deviations overlap substantially (e.g., Instance Precision 0.87±0.11 vs. CycleGAN 0.72±0.18; Panoptic Quality 0.65±0.10 vs. 0.59±0.14). The paper should report per-image paired statistics, bootstrap confidence intervals, or another appropriate significance test, or soften the 'significantly' language in the abstract, introduction, and conclusion.","section":"Introduction, Table 1, Conclusion"},{"comment":"The failure example demonstrates that when source and target domains do not share a common content space, the model hallucinates structures. This assumption is load-bearing for both applications, because the training signal for content preservation comes from cycle consistency across unpaired domains. The manuscript does not provide evidence that the real same-fluorophore breast/colon datasets satisfy this overlap condition beyond qualitative inspection, and the reader cannot assess how often artifacts similar to Fig. 14 occur in the intended acquisition regime. Please provide a quantitative characterization of content-space overlap for the IF data (e.g., artifact frequency, or detection of structure types absent from the other domain) or discuss the failure mode explicitly as a limitation in the main text.","section":"Appendix H"}],"minor_comments":[{"comment":"The model name is inconsistently capitalized: 'Ui2i' and 'UI2I' are both used (e.g., Sections 5 and 6); please standardize.","section":"Throughout"},{"comment":"Table 1 reports '0.72 ∓ 0.14' for Ui2i w/o augment. Instance Recall; this should be '±'.","section":"Table 1"},{"comment":"Equation (1) is typeset incorrectly in the provided PDF (the formula is garbled); please provide a clean rendering of the lower bound and verify the notation matches the cited source.","section":"Eq. (1)"},{"comment":"The text mentions that 'class embeddings are explained in the Appendix,' but Appendix A and B do not clearly define or illustrate these class embeddings for the three-class discriminator; please add a short explanation or remove the cross-reference.","section":"Section 6, Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The novelty claim is strong and will draw scrutiny from the microscopy and image-translation communities. The absence of a quantitative same-fluorophore validation is the main correctness risk; the HT-T24 surrogate is explicitly acknowledged in the manuscript as not reproducing the intended regime. The overlapping standard deviations in Table 1 also weaken the segmentation claim. I would recommend that the editor require the same-fluorophore quantitative evidence, or a clearly scoped revision of the novelty claim, before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Nothing here is going to change your life, but Ui2i is worth a look if you care about unpaired i2i in microscopy. The genuinely new piece is the task: separating two markers labeled with the same fluorophore from a single-channel image, trained on real unpaired data. That is not in the earlier literature; MicroSplit does semantic unmixing but needs paired aligned data. The architecture is a soup of known tricks—U-Net generators with skip connections, no feature normalization, bidirectional spectral norm, attention, contrastive loss, scale augmentation—but the combination is sensible and the paper gives a believable mechanical reason (Appendix E) why instance norm hurts content preservation.\n\nThe segmentation half is the stronger part. Evaluating IHC→H&E translation with an external pretrained StarDist is the right move. The ablations are there, the test set has 300 annotated patches, and beating InstanSeg, which was trained on IHC, is a real if modest result. But the word 'significantly' is doing work the data don't support: in Table 1 the standard deviations overlap on essentially every metric (e.g., panoptic quality 0.65±0.10 vs CycleGAN's 0.59±0.14). There is also a possible train/test leak: the test patches are randomly selected from the same Lyon19/EndoNuke datasets used for training, and the paper doesn't say the split is at whole-slide level. That needs to be clarified.\n\nThe unmixing section is where I part ways with the abstract. The central claim—first unpaired separation of same-fluorophore signals—is not actually tested on same-fluorophore data. The quantitative evaluation uses HT-T24, where SOX2 and Grasp65 are on spectrally distinct fluorophores, and the authors concede in Appendix G that this 'does not faithfully reproduce the signal characteristics, noise patterns, or artifact profiles of true single-fluorophore multiplexing.' The real breast/colon results are qualitative only. Given the paper's own Appendix H shows the model hallucinates structures when source content is missing from the target domain, the concern is concrete: if real single-channel images have bright artifacts not represented in the two-channel training set, the unmixing could be producing plausible-looking fictions. The theoretical point about infinitely many cycle-consistent solutions is correct, but the added losses don't get a direct test of whether they resolve that ambiguity for the same-fluorophore regime.\n\nSo: solid engineering paper about segmentation, interesting but unproven task claim about unmixing. The authors are honest about the limits, which I respect, but honesty isn't evidence. I'd send it to a serious referee—the task formulation and segmentation experiments justify that—but I'd want a same-fluorophore quantitative check (synthetic mixtures of real single-fluorophore images would do) and code/data release before accepting.","headline":"A well-engineered CycleGAN variant with a genuinely new task formulation, but the headline unmixing claim is only tested on a surrogate the authors admit isn't representative.","tokens_in":18358,"tokens_out":5444,"would_cite":false,"duration_ms":48386,"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":"Ui2i claims to be the first unpaired image-to-image translation model that separates superimposed fluorescence signals in real microscope images and adapts IHC images for nuclear segmentation without retraining the segmenter.","keywords":["unpaired image-to-image translation","cycle consistency","content preservation","spectral normalization","nuclear segmentation","domain adaptation","immunofluorescence unmixing","multiplexed immunofluorescence"],"falsifier":"A concrete test: acquire a sample with two markers truly labeled by the same fluorophore and independently verify their locations by imaging each marker in a separate round or with spectrally distinct fluorophores; apply Ui2i to the single-channel multiplexed image and compare the separated channels to the independently acquired ground truth. If per-channel PSNR and MicroMS-SSIM are no better than a trivial assignment, such as copying the input to both channels, the first-to-unmix-with-unpaired-data claim is falsified. A second, quicker check is to scan for hallucinated structures in the output that have no counterpart in the input, as in the Appendix H failure example.","tokens_in":17251,"feed_emoji":"🔬","tokens_out":5829,"duration_ms":57475,"temperature":0.7,"pith_summary":"Ui2i is an unpaired image-to-image translation model built for biomedical images, where preserving the shape and position of structures matters more than in natural-image translation. The paper's central claim is that Ui2i can translate IHC-stained tissue images into the H&E domain so that a segmenter trained only on H&E can then segment nuclei in IHC images without retraining, and that it can separate two markers captured together in one fluorescence channel into two clean channels. If true, the same method that adapts stains for segmentation also doubles the usable marker capacity of multiplexed immunofluorescence experiments, because one fluorophore can label two biomarkers and the model untangles them computationally. The paper reports that Ui2i preserves content fidelity to a significantly greater extent than existing translation methods, and that its unmixing of real single-channel images is the first to use unpaired real data rather than paired synthetic mixtures.","feed_headline":"Unpaired translation separates overlapping fluorescence signals","feed_subtitle":"The same model lifts IHC nuclear segmentation without retraining and lets one fluorophore label two biomarkers.","key_machinery":"The load-bearing mechanism is a set of architectural choices bundled under the name Ui2i: U-Net generators whose skip connections carry localized shallow features into the decoder, approximate bidirectional spectral normalization (a parameter-based normalization that replaces feature-based layers such as instance normalization and keeps local object responses independent of global context), channel-spatial attention in the encoders, and a shared bottleneck supervised by a content discriminator plus an N-pair contrastive loss. Cycle consistency ties the two translation directions together; the paper argues that cycle loss alone cannot pin down content preservation, so the extra bottlenecks and attention are what actually protect nuclear shapes and membrane patterns.","core_discovery":"At its core, Ui2i is CycleGAN rebuilt for content fidelity. It keeps the unpaired, bidirectional cycle-consistency setup but replaces ResNet generators with U-Net-style encoders and decoders linked by skip connections, removes feature-based normalization layers and instead normalizes weights with approximate bidirectional spectral normalization, adds channel and spatial attention in the encoders, and supervises the shared bottleneck with a content discriminator and a cross-domain contrastive loss. The paper demonstrates this design on two tasks: translating IHC images to H&E so that the H&E-pretrained StarDist segmenter outperforms the IHC-domain-pretrained InstanSeg on IHC test images, and unmixing single-channel immunofluorescence images of Ki67 and E-cadherin, labeled with the same fluorophore, into separate channels. The strongest claim, stated as a first in the field, is that this is the first approach able to separate superimposed signals in IF images using real, unpaired training data.","pith_inferences":["The same content-preservation machinery should transfer to other biomedical translation tasks where shape fidelity is critical, such as CT-to-MRI or stain transfer on whole-slide images; this is a natural extension the paper does not test.","The declared reliance on a shared content space implies a practical curation rule: any structure that appears in only one domain must be added to the other, or the model will hallucinate it; the failure example in Appendix H already shows this by adding artifact patches to the two-channel set.","If the doubling claim holds, multiplexed panels could be expanded combinatorially by co-labeling multiple marker pairs, not just one pair, though crosstalk between pairs would need separate validation.","A quantitative check of the shared-content assumption could be built by measuring how often the model's outputs contain structures that have no counterpart in the input, for example with an auxiliary detector trained on the target domain."],"forward_implications":["IHC-to-H&E translation by Ui2i followed by an H&E-pretrained StarDist segmenter yields higher instance precision and recall on IHC images than applying the IHC-pretrained InstanSeg directly, so domain-aligned translation can beat a generalist segmenter.","Single-channel IF unmixing lets one fluorophore label two biomarkers, effectively doubling multiplexed marker capacity and reducing the number of fluorophores and imaging rounds.","Because training is unpaired, the unmixing model can be built from real single-channel and two-channel acquisitions instead of synthetic paired mixtures, which the paper argues is closer to how co-labeled markers actually appear.","Removing feature-based normalization and using parameter-based spectral normalization suppresses blob-like hallucinations that would be counted as false-positive nuclei.","The pipeline needs no retraining of the downstream segmenter: the translation module alone adapts the input domain."],"supporting_citations":[{"why":"Supplies the base CycleGAN architecture and cycle-consistency adversarial training that Ui2i builds on.","marker":"[61]"},{"why":"Introduces spectral normalization, which Ui2i uses in place of feature-based normalization.","marker":"[38]"},{"why":"Defines bidirectional spectral normalization, the concept Ui2i approximates for convolutional layers.","marker":"[31]"},{"why":"Shows that instance normalization can cause blob-like artifacts and motivates parameter-based normalization.","marker":"[24]"},{"why":"Provides the H&E-pretrained StarDist segmenter used to evaluate translated IHC images.","marker":"[42]"},{"why":"Provides the IHC-pretrained InstanSeg baseline that Ui2i-based translation outperforms.","marker":"[8]"},{"why":"Describes the paired-data prior work for IF unmixing and supplies the HT-T24 dataset for quantitative comparison.","marker":"[3]"},{"why":"Defines the MicroMS-SSIM metric used to score the unmixing results.","marker":"[4]"},{"why":"Introduces the pixelwise contrastive loss that Ui2i adapts to its bottleneck features.","marker":"[40]"}],"fun_headline_variants":["Unpaired translation preserves content, separates IF signals","First unpaired method to unmix fluorescence channels","Ui2i: content-faithful unpaired translation for segmentation and unmixing","Unpaired cycleGAN keeps content, unmixes overlapping signals","Content-aware unpaired translation separates superimposed labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated in Appendix H, is that the source and target datasets share a common content space: if a structure appears in one domain and not the other, the model tends to hallucinate it.","fun_headline_variants_meta":{"raw":{"variants":["Unpaired translation preserves content, separates IF signals","First unpaired method to unmix fluorescence channels","Ui2i: content-faithful unpaired translation for segmentation and unmixing","Unpaired cycleGAN keeps content, unmixes overlapping signals","Content-aware unpaired translation separates superimposed labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000849,"raw_usage":{"total_tokens":3689,"prompt_tokens":934,"completion_tokens":2755,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":2676}},"tokens_in":550,"tokens_out":2755,"duration_ms":22671,"temperature":1.0,"reasoning_tokens":2676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:47:38.883396+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: acquire a sample with two markers truly labeled by the same fluorophore and independently verify their locations by imaging each marker in a separate round or with spectrally distinct fluorophores; apply Ui2i to the single-channel multiplexed image and compare the separated channels to the independently acquired ground truth. If per-channel PSNR and MicroMS-SSIM are no better than a trivial assignment, such as copying the input to both channels, the first-to-unmix-with-unpaired-data claim is falsified. A second, quicker check is to scan for hallucinated structures in the output that have no counterpart in the input, as in the Appendix H failure example.","supporting_citations":[{"cited_title":"Unpaired image-to- image translation using cycle-consistent adversarial networks","cited_arxiv_id":null,"evidence_quote":"Supplies the base CycleGAN architecture and cycle-consistency adversarial training that Ui2i builds on."},{"cited_title":"Spectral Normaliza- tion for Generative Adversarial Networks","cited_arxiv_id":null,"evidence_quote":"Introduces spectral normalization, which Ui2i uses in place of feature-based normalization."},{"cited_title":"Why Spectral Normalization Stabilizes GANs: Analysis and Improvements","cited_arxiv_id":null,"evidence_quote":"Defines bidirectional spectral normalization, the concept Ui2i approximates for convolutional layers."},{"cited_title":"Analyzing and Improving the Image Quality of StyleGAN","cited_arxiv_id":null,"evidence_quote":"Shows that instance normalization can cause blob-like artifacts and motivates parameter-based normalization."},{"cited_title":"Cell Detection with Star-Convex Polygons","cited_arxiv_id":null,"evidence_quote":"Provides the H&E-pretrained StarDist segmenter used to evaluate translated IHC images."},{"cited_title":"InstanSeg: an embedding-based instance segmentation algorithm optimized for accurate, efficient and portable cell segmentation, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the IHC-pretrained InstanSeg baseline that Ui2i-based translation outperforms."},{"cited_title":"MicroSplit: Semantic Unmixing of Fluorescent Microscopy Data","cited_arxiv_id":null,"evidence_quote":"Describes the paired-data prior work for IF unmixing and supplies the HT-T24 dataset for quantitative comparison."},{"cited_title":"Efros, Richard Zhang, and Jun-Yan Zhu","cited_arxiv_id":null,"evidence_quote":"Introduces the pixelwise contrastive loss that Ui2i adapts to its bottleneck features."}],"review_version":1}