{"id":"925879bb-7509-460b-8cc4-5576723e8df7","arxiv_id":"2504.13402","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CytoFM, a cytology-specific self-supervised foundation model, outperforms histopathology and natural-image models on two of three cytology classification tasks.","lead":"This paper introduces CytoFM, a self-supervised vision transformer pretrained on cytology images from breast, cervix, and thyroid samples. It outperforms non-cytology foundation models on two of three downstream cell classification tasks, including a fully unseen cervical dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MLBC/FNAC same-dataset splits permit patient/domain leakage; the only truly unseen win (HiCervix) lacks error bars, so 'two of three' is not yet established.","rationale":"I agree with the reader's weakest assumption. The key unresolved question is whether CytoFM's wins reflect cytology-specific representations or same-distribution familiarity. The paper has strengths: standard iBOT pretraining, a diverse pretraining corpus, and an honest discussion of the breast/cervix imbalance (Section 4). It also uses a fully unseen HiCervix set, which is the right idea, but the lack of uncertainty quantification there is a gap. A secondary concern is that UNI, pretrained at 20x, is evaluated on 40x patches; this may disadvantage UNI, but I do not make it the primary issue because the same-resolution protocol is at least consistent across models. The decisive fix is a patient-level split plus confidence intervals on the unseen-dataset result. Since the reader already assigned CONDITIONAL with moderate confidence, my read leaves that verdict unchanged.","tokens_in":9832,"tokens_out":6971,"duration_ms":66141,"concrete_test":"Recompute Table 2 for MLBC and FNAC2019 after splitting the data by patient ID, so that no patient contributes both pretraining and evaluation patches, and report bootstrap 95% confidence intervals for HiCervix accuracy and AUC (or multiple random splits). If CytoFM's MLBC advantage over UNI and iBOT-ImageNet disappears under a patient-level split, or if the HiCervix intervals overlap with the baselines, the 'two of three' headline should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing point is not that CytoFM underperforms, but that the empirical basis for 'two of three' is weaker than it appears. In §3.1 the two datasets that CytoFM wins on during partially held-out evaluation, MLBC and HiCervix, are not symmetric evidence: MLBC was also used in pretraining, with the held-out portion taken from the same dataset and, as described, split into 'two non-overlapping sets' at the image/patch level rather than at the patient level. MLBC contains 963 images from 460 patients, so a non-patient-level split can put images from the same patient or the same scanner/stain protocol into both pretraining and test. Self-supervised pretraining can memorize site-specific artifacts, so the MLBC advantage (0.930 vs 0.895 UNI and 0.879 iBOT-ImageNet) may be an artifact of distributional leakage rather than transferable cytology knowledge. The completely unseen HiCervix result is the only clean out-of-distribution win, but it is reported from a single fixed split without confidence intervals or a significance test, so it cannot by itself carry the 'two of three' claim. This combination is what makes the central claim conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript describes CytoFM, a ViT-Base encoder pretrained with iBOT on approximately 1.4 million image patches from seven public cytology datasets and one private thyroid dataset. The pretrained encoder is frozen and its features are aggregated with an attention-based multiple instance learning (ABMIL) framework for three downstream tasks: malignancy classification on FNAC2019 and cell-type classification on MLBC and HiCervix. The authors compare CytoFM against iBOT-ImageNet and UNI and report that CytoFM outperforms both baselines on MLBC and HiCervix but underperforms on FNAC2019. They also provide UMAP visualizations of feature embeddings and attention maps as qualitative evidence that CytoFM learns cytologically relevant features.","tokens_in":10115,"tokens_out":3993,"duration_ms":38190,"significance":"If the empirical claims hold, CytoFM would be a useful contribution as the first cytology-specific self-supervised foundation model. The comparison against a histopathology foundation model (UNI) and a natural-image model (iBOT-ImageNet) is appropriate and informative. Strengths of the paper include the use of multiple public datasets, the inclusion of a completely unseen evaluation dataset (HiCervix), significance testing over 100 random splits for two of the three tasks, and the qualitative UMAP and attention-map analyses. However, the evaluation protocol does not yet fully support the central 'two of three' generalization claim: the two partially held-out datasets use same-dataset image-level splits rather than patient-level splits, and the clean out-of-distribution result on HiCervix is reported from a single fixed split without confidence intervals or a significance test. These issues are fixable with additional experiments and statistical reporting, but they are load-bearing for the main claim.","major_comments":[{"comment":"The held-out portions of FNAC2019 and MLBC are formed by splitting the same datasets used for pretraining into 'two non-overlapping sets' at the image or patch level, not at the patient level. MLBC is described as containing 963 images from 460 patients, so the current split can place images from the same patient or the same scanner/stain protocol into both the pretraining set and the test set. Self-supervised pretraining can memorize site-specific acquisition artifacts, which would inflate CytoFM's performance on MLBC (accuracy 0.930 versus 0.895 for UNI and 0.879 for iBOT-ImageNet) relative to an unseen dataset. This is the load-bearing weakness identified by the stress-test concern, and it does land: the MLBC advantage cannot be cleanly attributed to transferable cytology knowledge without a patient-level split or an equivalent control. Please report results with patient-level splits for FNAC2019 and MLBC, or otherwise demonstrate that the image-level split does not create distributional overlap.","section":"§3.1, Table 2"},{"comment":"The HiCervix result is the cleanest out-of-distribution evidence for the generalization claim, yet it is reported from a single fixed train/validation/test split with no confidence intervals, error bars, or significance test. The observed accuracy gap (0.844 for CytoFM versus 0.803 for iBOT-ImageNet and 0.800 for UNI) is promising, but it cannot by itself carry the 'two of three' claim. Please add bootstrap confidence intervals or repeated-seed variability estimates and perform a statistical comparison against the baselines on the same fixed split. Without this, the HiCervix gain is not quantitatively established.","section":"§3.2, Table 2"},{"comment":"The private thyroid dataset is filtered with the custom ThyVGG model using a top-1500 patch threshold, and the paper acknowledges in §4 that breast cytology contributes only 2,544 images while cervical cytology contributes close to a million. Since no downstream thyroid evaluation or ablation is presented, the claim that CytoFM learns representations generalizing across organs is not currently supported by the experiments. Either add a thyroid evaluation or soften the cross-organ generalization claim in the introduction and discussion.","section":"§2.2, Figure 2; §4"}],"minor_comments":[{"comment":"The spelling 'Papanicolau' should be 'Papanicolaou' in the description of the FNAC2019 dataset.","section":"§2.1"},{"comment":"The method name 'HeirSwin' appears to be a typo for 'HierSwin'; please correct it and ensure the reference matches the cited work.","section":"§4"},{"comment":"The attention maps in Figure 3 are selected examples and may not be representative; a quantitative analysis, such as attention-agreement with cell-level annotations, would strengthen the claim that CytoFM attends to cytologically relevant features.","section":"§3.2, Figure 3"},{"comment":"The manuscript does not mention a plan to release code, model weights, or the private thyroid dataset; given the emphasis on reproducibility and downstream reuse, a statement about data and code availability would be helpful.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sensible and the empirical setup is on the right track, but the evaluation protocol needs strengthening before the central generalization claim is publication-ready. The 'first cytology foundation model' claim should also be checked against concurrent preprints and recent work, as novelty screening may turn up other cytology-specific self-supervised models. I recommend major revision rather than rejection because the identified issues are addressable with additional experiments and statistical reporting, not by a fundamentally different approach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first self-supervised cytology foundation model I know of, and the authors deserve credit for assembling a multi-institutional pretraining corpus and showing that domain-specific pretraining helps in at least one clean out-of-distribution setting. But the evidence for \"two of three\" is not as strong as the abstract implies. The MLBC win is confounded by non-patient-level splitting, and the HiCervix win is reported without error bars.\n\nThe genuinely new thing is empirical: they train an iBOT ViT-Base on ~1.4M patches from eight cytology datasets spanning breast, cervix, and thyroid, and compare frozen features via ABMIL against iBOT-ImageNet and UNI. The HiCervix result — accuracy 0.844 vs 0.803/0.800, AUC 0.968 vs 0.956/0.952 — is the only clean generalization test, because that dataset is completely unseen. That result is a real positive, even if modest.\n\nThe soft spots are in the evaluation protocol. FNAC2019 and MLBC are split into \"non-overlapping sets\" for pretraining and evaluation, but MLBC has 963 images from 460 patients and the split is described at the image/patch level, not patient level. So the same patient's cells can appear in both pretraining and test, and site-specific artifacts (scanner, stain) can leak. The MLBC accuracy gain (0.930 vs 0.895/0.879) could be partly distributional. The authors do report significance tests for MLBC, which helps, but it does not remove the confound. HiCervix is the clean win, but with a single fixed split and no confidence intervals, one cannot tell if 0.844 vs 0.803 is stable. That is a minor-to-moderate issue, not fatal, because the direction is consistent with the MLBC result and with the UMAP/attention visualizations.\n\nTwo smaller things: no code or model release, which is becoming the norm for foundation models and matters for reproducibility; and the private thyroid dataset relies on a custom ThyVGG filter with a top-1500 threshold, a reasonable heuristic but another free parameter. The citation pattern is fine, and the discussion honestly flags the breast/cervix imbalance and the comparison to HiCervix's HeirSwin state of the art.\n\nBottom line: this is a paper I would send to a serious referee. The central claim that cytology-specific self-supervised pretraining helps is plausible and partially supported; the protocol needs patient-level splits and error bars on HiCervix. I would not desk reject it, but I would expect major revisions.","headline":"A genuinely first cytology foundation model with a plausible transfer signal, but the same-dataset evaluation splits and missing error bars on the unseen test make the 'two of three' claim weaker than the abstract suggests.","tokens_in":10614,"tokens_out":2123,"would_cite":false,"duration_ms":18558,"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":"CytoFM, the first cytology-specific foundation model, learns transferable cell representations that beat non-cytology foundation models on two of three downstream cytology classification tasks, including a dataset completely unseen during…","keywords":["cytology foundation model","self-supervised pretraining","iBOT","vision transformer","multiple instance learning","cervical cytology","breast cytology","cell morphology"],"falsifier":"A direct test: retrain CytoFM with all MLBC and FNAC2019 images removed from the pretraining corpus and rerun the same downstream protocols. If the reported MLBC and FNAC2019 margins over UNI and iBOT-ImageNet shrink or disappear, then the apparent advantage comes from training/test distribution overlap rather than from cytology-specific representation learning; the HiCervix result alone would then be the only clean transfer evidence.","tokens_in":9708,"feed_emoji":"🔬","tokens_out":8375,"duration_ms":70170,"temperature":0.7,"pith_summary":"CytoFM is presented as the first foundation model pretrained specifically on cytology images rather than on natural images or tissue slides. The paper's claim is that a Vision Transformer trained without labels—by a combination of masked-image prediction and self-distillation—on roughly 1.4 million cell patches from breast, cervix, and thyroid learns transferable representations of cell morphology. On two of three downstream classification tasks, including a cervical dataset the model never saw, those representations beat a histopathology foundation model and an ImageNet-trained model; on the third task, breast malignancy classification, CytoFM trails the ImageNet model. If the claim holds, cytology would gain a reusable feature extractor for diagnostic tasks that each have too few labels to train a model from scratch.","feed_headline":"First cytology-trained model beats general vision models on cell typing","feed_subtitle":"Pretrained on 1.4M cell patches, CytoFM also leads on a cervical dataset it never saw during training.","key_machinery":"The load-bearing object is CytoFM, a ViT-Base feature extractor pretrained with iBOT. iBOT trains a teacher-student pair with two objectives: self-distillation, where the student's class token must match the teacher's predictions on a differently augmented view, and masked image modeling, where the student predicts patch tokens for masked regions from the teacher's unmasked embeddings. This pairing lets one frozen encoder represent both coarse context, such as cellularity and cell clusters, and fine detail, such as nuclear boundaries and mitotic activity. For evaluation, patch features are pooled by an attention-based multiple instance learning head, so predictions are made at the image or slide level without patch-level labels.","core_discovery":"CytoFM is a frozen ViT-Base encoder obtained by continuing iBOT pretraining on a multi-institutional, multi-organ cytology corpus. On a held-out split of the MLBC cervical dataset, it reaches 0.930 ± 0.05 accuracy and 0.993 ± 0.01 AUROC (area under the ROC curve), significantly above UNI (0.895/0.986) and iBOT-ImageNet (0.879/0.983). On the completely unseen HiCervix dataset, it reaches 0.844 accuracy and 0.968 AUROC, ahead of UNI (0.800/0.952) and iBOT-ImageNet (0.803/0.956). On FNAC2019 breast malignancy classification it scores 0.908 accuracy and 0.979 AUROC, the lowest of the three models, which the authors attribute to breast cytology being only a small fraction of the pretraining data. The paper reads these results, together with attention maps that highlight nuclei, mitotic figures, and cell boundaries, as evidence that cytology-specific self-supervised pretraining captures features that general-purpose encoders miss.","pith_inferences":["A stronger test of the 'cytology-specific' claim would evaluate CytoFM on an organ absent from pretraining, such as urine or lung cytology; HiCervix is unseen but still cervical, so it mainly proves cross-institution transfer within the dominant organ.","The two partially held-out evaluations could be rerun with pretraining ablations that remove MLBC and FNAC2019 entirely, separating the effects of dataset overlap from genuine transfer.","If the HiCervix margin holds under such ablations, cytology foundation models could become the default frozen backbones for rare cytology tasks where labels are too scarce to train a supervised model from scratch.","The attention maps point to a concrete clinical extension the paper does not pursue: using CytoFM's attention as a screening signal for mitotic activity, which could be validated against expert annotations."],"forward_implications":["CytoFM's frozen features place it ahead of both baselines on cervical cell typing (MLBC accuracy 0.930, AUROC 0.993), so practitioners using a cytology-specific backbone may not need to train separate encoders for each cervical cytology task.","On a dataset the model never saw (HiCervix), CytoFM still leads (accuracy 0.844, AUROC 0.968), suggesting the learned features transfer across institutions and scanners without stain normalization or fine-tuning.","On breast FNAC2019, where breast patches were only a small part of pretraining, CytoFM trails iBOT-ImageNet; the paper's explanation implies that balancing organ representation in pretraining should improve cross-organ performance.","The reported gains justify further investment in larger, more balanced cytology pretraining corpora and in probing which cytology-specific features, such as mitotic activity and nuclear boundaries, drive the downstream improvements."],"supporting_citations":[{"why":"Supplies the iBOT self-supervised pretraining framework combining masked image modeling and self-distillation that trains CytoFM.","marker":"[25]"},{"why":"Provides the UNI histopathology foundation model used as the main pathology-specific baseline and motivates patch-based pretraining choices.","marker":"[4]"},{"why":"Provides the FNAC2019 breast cytology dataset used both in pretraining and in held-out malignancy classification evaluation.","marker":"[20]"},{"why":"Provides the MLBC cervical cytology dataset used both in pretraining and in held-out cell type classification evaluation.","marker":"[10]"},{"why":"Provides the HiCervix cervical dataset, completely unseen during pretraining, that tests cross-institution generalization.","marker":"[2]"},{"why":"Documents that histopathology-trained foundation models struggle on cytology, motivating the need for cytology-specific pretraining.","marker":"[5]"},{"why":"Grounds the weakly supervised multiple instance learning approach used to aggregate patch features for slide-level classification.","marker":"[6]"}],"fun_headline_variants":["CytoFM, first cytology foundation model, beats UNI and iBOT on 2 of 3 tasks","CytoFM tops cervical cancer benchmarks, falls short on breast malignancy","First cytology-specific foundation model outperforms generic AI on cervical tasks","CytoFM: cytology-trained model leads on two of three downstream tasks","New cytology AI foundation model beats histopathology and image models on cervical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that splitting the same dataset into disjoint pretraining and test portions measures generalization; because the pretraining and held-out FNAC2019 and MLBC images come from the same acquisition and labeling process, shared staining and scanner characteristics could inflate CytoFM's scores on those two tasks.","fun_headline_variants_meta":{"raw":{"variants":["CytoFM, first cytology foundation model, beats UNI and iBOT on 2 of 3 tasks","CytoFM tops cervical cancer benchmarks, falls short on breast malignancy","First cytology-specific foundation model outperforms generic AI on cervical tasks","CytoFM: cytology-trained model leads on two of three downstream tasks","New cytology AI foundation model beats histopathology and image models on cervical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000773,"raw_usage":{"total_tokens":3473,"prompt_tokens":1050,"completion_tokens":2423,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":2318}},"tokens_in":666,"tokens_out":2423,"duration_ms":14717,"temperature":1.0,"reasoning_tokens":2318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:08:23.296251+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: retrain CytoFM with all MLBC and FNAC2019 images removed from the pretraining corpus and rerun the same downstream protocols. If the reported MLBC and FNAC2019 margins over UNI and iBOT-ImageNet shrink or disappear, then the apparent advantage comes from training/test distribution overlap rather than from cytology-specific representation learning; the HiCervix result alone would then be the only clean transfer evidence.","supporting_citations":[{"cited_title":"Chen, Tong Ding, Ming Y","cited_arxiv_id":null,"evidence_quote":"Provides the UNI histopathology foundation model used as the main pathology-specific baseline and motivates patch-based pretraining choices."},{"cited_title":"Mahanta, and Anup Kumar Das","cited_arxiv_id":null,"evidence_quote":"Provides the FNAC2019 breast cytology dataset used both in pretraining and in held-out malignancy classification evaluation."},{"cited_title":"Mahanta, Himakshi Borah, and Chandana Ray Das","cited_arxiv_id":null,"evidence_quote":"Provides the MLBC cervical cytology dataset used both in pretraining and in held-out cell type classification evaluation."},{"cited_title":"HiCervix: An Extensive Hierarchical Dataset and Benchmark for Cervi- cal Cytology Classification","cited_arxiv_id":null,"evidence_quote":"Provides the HiCervix cervical dataset, completely unseen during pretraining, that tests cross-institution generalization."},{"cited_title":"Kovalsky, Serge Assaad, Jonathan Cohen, Danielle Elliott Range, Avani A","cited_arxiv_id":null,"evidence_quote":"Grounds the weakly supervised multiple instance learning approach used to aggregate patch features for slide-level classification."}],"review_version":1}