{"id":"65e55d6a-f215-4de4-9af0-a847754f8386","arxiv_id":"2412.02858","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Unpaired image translation between labeled and unlabeled microscopy domains can produce pseudo labels useful for segmentation, achieving 0.736 mean Dice for axons on SEM via the tutoring path.","lead":"This paper tests two ways to generate pseudo labels for histology image segmentation by translating between a labeled microscopy modality (TEM) and unlabeled target modalities (TEM-MACAQUE, SEM, bright-field). The tutoring path, which trains a segmenter on translated synthetic images, reaches a mean axon Dice of 0.736 on SEM, suggesting a useful starting point for manual annotation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.736 Dice result rests on a translation checkpoint selected by cycle-reconstruction metrics, not by semantic preservation; whether this checkpoint is representative of the pipeline remains untested.","rationale":"The paper presents a practical and clearly evaluated pseudo-labeling pipeline, and the reported 0.736 +/- 0.005 Dice on SEM is a genuine empirical result computed against real ground truth. That evaluation partially supports the claim for the particular checkpoint used. However, the reader's weakest assumption about semantic preservation is real and, in my view, the most load-bearing gap. The selection of the translation checkpoint is made with cycle-reconstruction metrics and visual inspection, neither of which directly measures whether the translated images retain the structures needed for the TEM annotations to transfer. The downstream Dice result answers whether this one checkpoint worked, but it does not answer whether the method, as described, reliably produces such checkpoints. Because the paper's stated contribution is a general pseudo-labeling pipeline and the tutoring path's headline number is a single checkpoint result, the absence of any downstream evaluation across checkpoints leaves the central claim conditional. This does not require changing the reader's CONDITIONAL verdict, but it sharpens the condition: the authors should demonstrate that checkpoint selection by cycle metrics is predictive of segmentation utility, or at least that the reported Dice is not a selected outlier.","tokens_in":7465,"tokens_out":5960,"duration_ms":99977,"concrete_test":"Retrain the tutoring segmentation model on synthetic TEM->SEM images generated from at least three to five SynDiff checkpoints bracketing the selected epoch (e.g., 40, 45, 50, 55, 60), keeping the nnU-Net configuration fixed, and compute SEM Dice for each checkpoint. Then measure the Spearman correlation between the cycle metrics (SSIM, PSNR, L1) and downstream Dice across these checkpoints. If Dice varies substantially across checkpoints or does not correlate with the selection metrics, the checkpoint-selection protocol is not a reliable proxy and the single reported 0.736 is not established as a robust central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is an empirical Dice value, but the pathway to that value depends on choosing one SynDiff checkpoint for the TEM-to-SEM translation. Section 3 states that checkpoints were selected using SSIM, PSNR, and L1 after an L->U->L reconstruction cycle, plus visual inspection. These metrics measure low-level fidelity to the source image, not whether the intermediate L->U translation preserves axon and myelin geometry, boundaries, or relative layout. If the selected checkpoint happened to preserve semantics while neighboring checkpoints do not, the reported 0.736 +/- 0.005 would be an artifact of checkpoint selection rather than a stable property of the tutoring pipeline. The downstream Dice on real SEM is direct evidence for the chosen checkpoint, but it does not validate the selection rule itself, and no non-selected checkpoint was evaluated downstream. Thus the central claim is load-bearing on an untested assumption: checkpoint quality measured by cycle reconstruction correlates with downstream segmentation utility.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an unsupervised pseudo-labeling pipeline for axon/myelin segmentation in histology images. It uses SynDiff, a cycle-consistent adversarial diffusion model, to translate between an annotated TEM dataset and three unannotated target datasets (TEM-MACAQUE, SEM, and BF). Two pseudo-labeling strategies are evaluated: tutoring, where labeled TEM images are translated into the target modality and an nnU-Net is trained on the resulting synthetic images with the original labels; and adaptation, where target images are translated into the TEM style and segmented with a pre-trained TEM model. Experiments against real target-domain ground truth show that tutoring achieves a mean Dice of 0.736 ± 0.005 on SEM, while the two strategies perform comparably on TEM-MACAQUE and adaptation performs better on BF. The authors conclude that the pipeline provides useful initialization masks for manual annotation.","tokens_in":7648,"tokens_out":7734,"duration_ms":76156,"significance":"If the reported results are robust, the paper offers a practical way to reuse annotated TEM data for SEM and BF, where large-scale annotation is costly. The evaluation is carried out against real ground truth in the target domains, uses 5-fold cross-validation with reported standard deviations, and the code is public. The comparison of two pseudo-labeling strategies across three target datasets with increasing domain shift is a useful empirical contribution. Importantly, there is no circularity: target labels are used only for evaluation, not for training the translation model or for selecting checkpoints. The main caveats are that the headline Dice depends on a single translation checkpoint selected by reconstruction metrics, and that no alternative pseudo-labeling or domain-adaptation baseline is tested; these limit the strength of the claim that this specific pipeline is responsible for the observed result.","major_comments":[{"comment":"The headline result of 0.736 ± 0.005 for the SEM tutoring path is obtained with a single SynDiff checkpoint selected by cycle-reconstruction SSIM/PSNR/L1 and visual inspection. These metrics measure low-level fidelity to the source rather than preservation of axon and myelin geometry, so the reported Dice may be specific to the chosen epoch (50) rather than a stable property of the pipeline. I request an analysis of downstream Dice across multiple translation checkpoints, or at least a comparison between the selected checkpoint and its neighbors, to establish that the checkpoint-selection rule does not inadvertently favor a particularly good segmentation outcome.","section":"Section 3, Fig. 3"},{"comment":"The only baseline for the two large-domain-shift experiments is a pre-trained TEM model that produces empty masks. To support the claim that the proposed SynDiff-based pseudo-labeling pipeline is responsible for the observed Dice scores, the paper should compare against at least one alternative unsupervised translation or domain-adaptation method, such as a CycleGAN-based version of the same tutoring protocol or a standard unsupervised domain adaptation baseline. Without such a comparison, the central result is not contextualized, and the specific contribution of the proposed pipeline remains unclear.","section":"Section 2.3, Table 2"},{"comment":"The recommendation to \"first use the adaptive path\" is inconsistent with the SEM results, where the adaptive path yields a mean Dice of 0.238 ± 0.088 for axons while the tutoring path yields 0.736 ± 0.005. The recommendation should be made conditional on the target domain, or the paper should provide a concrete criterion (for example, a small labeled validation set) for deciding between the two paths.","section":"Conclusion vs. Table 2"}],"minor_comments":[{"comment":"The mechanism by which the nnU-Net softmax outputs are converted to final pseudo-label masks (threshold value, post-processing, connected components) is not specified; providing this detail would improve reproducibility.","section":"Section 2.3"},{"comment":"It is not stated whether the SynDiff translation model is trained on the same XU images that are later used for evaluation; if so, the evaluation is transductive. This is acceptable for a pseudo-labeling pipeline, but it should be disclosed explicitly because it affects claims about applicability to new unlabeled datasets.","section":"Section 2.2/2.3"},{"comment":"The claim that a Dice score above 0.5 reduces annotation time by 25-50% is unsupported by any data or citation in the manuscript; since the authors themselves call for future quantification, this sentence should be moved to future work or supported with evidence.","section":"Conclusion"},{"comment":"The standard deviation for the SEM adaptation path is very high (0.088 for axon Dice), and the low sensitivity (0.144) is mentioned in the text, but a brief explanation of the failure mode (for example, near-empty predicted masks) would help the reader interpret this result.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid applied contribution with an honest evaluation design, and the code availability is appreciated. The main risk is the checkpoint-selection robustness: the headline Dice rests on a single SynDiff epoch chosen by reconstruction metrics. If the authors can provide a checkpoint-sensitivity analysis and at least one alternative baseline, I would be comfortable with acceptance. The missing baseline comparison is a common limitation in applied papers, but in this case it weakens the causal claim about the specific pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a clean empirical paper that extends translation-based pseudo labeling from detection to segmentation, with code and real ground-truth evaluation. The headline number (0.736 Dice on SEM via the tutoring path) looks real, but the paper doesn't show it's robust to the single checkpoint choice, and the conclusion recommends a strategy that their own table contradicts.\n\nWhat's genuinely useful: the tutoring-vs-adaptation comparison across three target domains is a simple, well-described experiment that gives practitioners a practical rule of thumb. The evaluation uses real labels, 5-fold CV with standard deviations, and the code is public. The TEM-MACAQUE result nicely shows when the pipeline is unnecessary, and the BF result shows a case where adaptation beats tutoring. That is a useful map of behavior.\n\nThe soft spot: checkpoint selection. Section 3 says they picked SynDiff epochs by SSIM/PSNR/L1 on the L->U->L cycle plus visual inspection, not by downstream segmentation quality. So the 0.736 ± 0.005 could be an artifact of picking a lucky epoch. The stress-test note has it right: the paper never tests whether neighboring checkpoints give similar Dice. This is addressable: report Dice across epochs or at least show that the selected checkpoint is on a plateau. Without that, the title-level claim is conditional.\n\nSecond, the baseline is only a pre-trained TEM model that returns empty masks on SEM/BF. No comparison to standard pseudo-labeling (self-training, confidence filtering) or to a CycleGAN-based translation. That makes it hard to know how much SynDiff matters.\n\nThird, the conclusion recommends \"first use the adaptive path,\" but on SEM the tutoring path is three times better in Dice (0.736 vs 0.238). That is not a minor wording issue; it will mislead readers. They need to state that adaptive is only preferred when compute is limited or the translation direction is easier.\n\nThe annotation-time savings (25-50%) is flagged as speculative in the text, so I don't hold it against them.\n\nWho this is for: groups doing annotation-acceleration for microscopy, especially histology; they will get a concrete workflow and baseline numbers. It deserves a serious referee, but the revision should include checkpoint robustness, at least one stronger baseline, and a fixed recommendation.\n\nI'd send it to review with major-revision expectations.","headline":"Useful practical pseudo-labeling pipeline with an untested checkpoint-selection assumption and a recommendation that contradicts its own SEM results.","tokens_in":8138,"tokens_out":4886,"would_cite":false,"duration_ms":45767,"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":"The paper establishes that translating labeled TEM histology images into SEM style and training a segmentation model on the synthetic images yields pseudo labels with mean Dice $0.736 \\pm 0.005$, a usable starting point for manual…","keywords":["pseudo labeling","unsupervised image translation","histology segmentation","axon and myelin segmentation","cycle-consistent diffusion","scanning electron microscopy","transmission electron microscopy","nnU-Net"],"falsifier":"Take a set of real SEM images with expert axon and myelin annotations, run the tutoring path using checkpoints selected by cycle-reconstruction SSIM, PSNR, and L1 exactly as the paper describes, and compare the pseudo labels to the expert masks on the same images. If the Dice score does not replicate near $0.736$ for axons, or if errors concentrate at myelin boundaries, the claim that cycle-consistent translation preserves geometry for labeling is weakened. A sharper variant is to re-select the translation checkpoint by downstream segmentation Dice instead of image-quality metrics; if Dice changes materially, the paper's checkpoint rule is not the one that best serves pseudo labeling.","tokens_in":7280,"feed_emoji":"🔬","tokens_out":8220,"duration_ms":74685,"temperature":0.7,"pith_summary":"Histology segmentation is bottlenecked by expert annotation, and annotations made in one imaging modality such as transmission electron microscopy (TEM) do not transfer to another such as scanning electron microscopy (SEM). This paper proposes to bridge the gap with unsupervised image translation: translate labeled images into the target modality, keep the old annotations on the translated images, and train a segmentation model on the synthetic pairs. The headline result is that this tutoring path produces pseudo labels on a real SEM dataset with a mean Dice score of $0.736 \\pm 0.005$ for axons and $0.652 \\pm 0.005$ for myelin. If correct, the method would let researchers obtain initial segmentations for a new modality with zero target-domain annotations, reducing the manual refinement burden.","feed_headline":"TEM-to-SEM translation yields pseudo labels at 0.736 Dice","feed_subtitle":"Tutoring path trains on synthetic SEM images, producing axon and myelin masks ready for manual refinement.","key_machinery":"The central mechanism is a cycle-consistent adversarial diffusion translation model (SynDiff) paired with an automatic segmentation framework (nnU-Net). SynDiff learns unpaired translations between the labeled domain $L$ and the unlabeled domain $U$ in both directions: a pair of non-diffusive GANs produce initial coarse translations, and denoising diffusion GANs refine them in fewer than eight timesteps, with a cycle-consistency loss forcing $L \\to U \\to L$ and $U \\to L \\to U$ to reconstruct the original images. The tutoring path takes the $L \\to U$ translation, recycles the original annotations as if they labeled the synthetic images, trains nnU-Net on those synthetic pairs, and applies it to real $U$ images. The adaptive path takes the $U \\to L$ translation and runs a model pre-trained on $L$ on the translated images. What carries the argument is that cycle consistency is assumed to preserve the geometry and boundaries of axons and myelin, so labels remain meaningful across a modality change.","core_discovery":"The paper claims that unpaired modality translation can generate usable pseudo labels for histology segmentation in a target domain that has no annotations. Using a labeled TEM dataset as the source, it translates images into three increasingly different target domains (TEM-MACAQUE, SEM, and BF) and tests two labeling strategies. The tutoring path, which translates labeled TEM images into the target modality, trains an nnU-Net on the synthetic images with the recycled annotations, and applies it to real target images, achieves a mean Dice score of $0.736 \\pm 0.005$ for axons and $0.652 \\pm 0.005$ for myelin on the SEM dataset. The adaptive path, which translates unlabeled target images into TEM style and applies a pre-trained TEM model, lags on SEM with axon Dice $0.238 \\pm 0.088$ but surpasses tutoring on the BF dataset with axon Dice $0.586 \\pm 0.025$ versus $0.454 \\pm 0.010$. The paper concludes that when the domain shift is large enough to defeat a pre-trained model, translation-based pseudo labeling provides useful initial masks for manual correction.","pith_inferences":["Editorial extension: if translation preserves geometry, the same pipeline should bootstrap segmentation for other unpaired microscopy pairs (bright-field to CARS, different stains), since nothing in the method depends on electron microscopy specifically.","Editorial extension: selecting translation checkpoints by downstream segmentation Dice on a tiny manually labeled subset could outperform the cycle-reconstruction metrics (SSIM, PSNR, L1) used in the paper, and would directly test whether reconstruction quality is the right proxy for label utility.","Editorial extension: the claimed 25–50 percent annotation-time saving at Dice above 0.5 is plausible but unmeasured; a timing study of manual correction of pseudo labels versus manual annotation from scratch would give the practical benefit real numbers."],"forward_implications":["On the SEM dataset, tutoring-path pseudo labels reach $0.736 \\pm 0.005$ Dice for axons and $0.652 \\pm 0.005$ for myelin, high enough to serve as initialization for manual refinement.","When the labeled and unlabeled domains are close (TEM versus TEM-MACAQUE), pseudo labeling adds nothing over applying the pre-trained model directly, so translation-based pseudo labeling is only worth the compute when the domain shift is large.","On the larger BF domain shift, the adaptive path outperforms tutoring, with axon Dice $0.586 \\pm 0.025$ versus $0.454 \\pm 0.010$; the best strategy therefore depends on translation quality in each direction.","The recommended workflow is to run the adaptive path first because it needs no extra segmentation model, and to use the tutoring path to generate a complementary set of pseudo labels at the cost of training a proxy model."],"supporting_citations":[{"why":"Supplies SynDiff, the adversarial diffusion translation model used to translate between labeled and unlabeled domains.","marker":"[10]"},{"why":"Provides the denoising diffusion GAN formulation that SynDiff uses for few-timestep sampling.","marker":"[17]"},{"why":"Supplies the pix2pix ResNet generator and discriminator architecture used for the non-diffusive cycle-consistency modules.","marker":"[7]"},{"why":"Supplies the nnU-Net self-configuring segmentation framework used to train proxy models on synthetic translated data.","marker":"[5]"},{"why":"Supplies the nnU-Net validation protocol that motivates the 5-fold cross-validation used for Dice scores.","marker":"[6]"},{"why":"Earlier cross-modality pseudo labeling for microscopy image quantification; the paper contrasts its semantic segmentation approach with this predecessor.","marker":"[18]"}],"fun_headline_variants":["Pseudo labeling via TEM-to-SEM translation hits 0.736 Dice","Unpaired translation generates histology pseudo labels at 0.736 Dice","Tutoring path: translate TEM to SEM for axon masks at 0.736 Dice","No target labels needed: pseudo labels from unpaired translation reach 0.736","TEM-SEM translation crafts pseudo labels for histology segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes the TEM-to-SEM translation preserves axon and myelin geometry and boundaries, so annotations made on original TEM images remain valid for translated images and transfer to real SEM.","fun_headline_variants_meta":{"raw":{"variants":["Pseudo labeling via TEM-to-SEM translation hits 0.736 Dice","Unpaired translation generates histology pseudo labels at 0.736 Dice","Tutoring path: translate TEM to SEM for axon masks at 0.736 Dice","No target labels needed: pseudo labels from unpaired translation reach 0.736","TEM-SEM translation crafts pseudo labels for histology segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1318,"prompt_tokens":912,"completion_tokens":406,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":306}},"tokens_in":528,"tokens_out":406,"duration_ms":4166,"temperature":1.0,"reasoning_tokens":306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:00:20.358130+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of real SEM images with expert axon and myelin annotations, run the tutoring path using checkpoints selected by cycle-reconstruction SSIM, PSNR, and L1 exactly as the paper describes, and compare the pseudo labels to the expert masks on the same images. If the Dice score does not replicate near $0.736$ for axons, or if errors concentrate at myelin boundaries, the claim that cycle-consistent translation preserves geometry for labeling is weakened. A sharper variant is to re-select the translation checkpoint by downstream segmentation Dice instead of image-quality metrics; if Dice changes materially, the paper's checkpoint rule is not the one that best serves pseudo labeling.","supporting_citations":[{"cited_title":"IEEE Transactions on Medical Imaging (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies SynDiff, the adversarial diffusion translation model used to translate between labeled and unlabeled domains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier cross-modality pseudo labeling for microscopy image quantification; the paper contrasts its semantic segmentation approach with this predecessor."}],"review_version":1}