{"id":"37df1458-8b63-467d-80c9-dc7601cb13ca","arxiv_id":"2509.03614","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A combined segmentation and classification teacher-student model scored F1 0.7660 for mitosis detection and balanced accuracy 0.8418 for atypical mitosis classification on the MIDOG 2025 preliminary test set.","lead":"The authors describe a teacher-student deep learning system that finds and classifies dividing cells (mitoses) in stained tumor images, and report how it scored on the MIDOG 2025 challenge's preliminary test set. If the scores hold, it is a workable approach to automating part of tumor grading, with methods aimed at robustness across stains and tissue domains.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation or baseline means the claimed effectiveness of pseudo-mask supervision is unsupported; morphological pseudo-masks may even label mitotic figures as normal nuclei.","rationale":"The reader identifies the reliability of teacher-generated pseudo-masks as the weakest assumption and returns a CONDITIONAL verdict. I agree that this is a central weakness, but I would sharpen it: the evidence does not even establish that the pseudo-masks help, because no ablation or baseline is reported. The manuscript explicitly admits the lack of ablation, which is a missing support that must be flagged. The specific failure mode—morphological pseudo-masks labeling mitotic figures as normal nuclei—is concrete and testable, and it directly threatens the central effectiveness claim. This is not an attack on the reported numbers; the challenge test scores may well be accurate. But the paper's interpretation of those scores as evidence for the proposed teacher-student mechanism is unjustified. The inconsistency between the abstract's 0.8414 and the results' 0.8418 is minor and does not change the verdict. Since the reader already set the verdict to CONDITIONAL on similar grounds, my read does not move the verdict. I recommend the authors provide the missing ablations, especially a pseudo-mask-on/off comparison, before the effectiveness claim can be accepted.","tokens_in":6891,"tokens_out":5954,"duration_ms":66654,"concrete_test":"Run a controlled ablation on the same data splits used for Table 4: (a) full model; (b) identical pipeline but with the normal-nucleus pseudo-mask term removed from Lsemi, keeping all other losses, augmentations, and threshold selection identical; (c) plain UNet without DG modules and without pseudo-masks. Evaluate all three on the same in-domain test set and, if accessible, the preliminary test set. If (b) and (c) achieve BA within 0.01 of the full model, the claimed contribution is not supported; if (b) is substantially worse, the pseudo-mask concern is resolved. Additionally, compute pixel overlap between morphological pseudo-masks and annotated mitosis masks; if more than 10% of mitosis pixels are covered by 'normal nuclei' pseudo-mask labels, the conflict is real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's core claim—'demonstrating the effectiveness of integrating segmentation-based detection and classification into a unified framework'—is not supported by the experiments as reported. The paper's own Discussion states: 'this study is limited by the absence of a detailed ablation analysis to quantify the contribution of each component.' Without a baseline (e.g., plain UNet without pseudo-masks, contrastive, DANN, or multi-scale classifier), the reported F1=0.7660 and BA=0.8418 cannot be attributed to any specific component. This is especially acute for Track 2, where the 'teacher' pseudo-masks are not produced by a learned teacher but by classical morphology: H-channel extraction, Gaussian blur, adaptive Otsu thresholding, and morphological opening. Mitotic figures are also hematoxylin-dense and will often fall inside the thresholded regions, so the pseudo-mask can label mitotic pixels as 'normal nuclei.' The manuscript never states that annotated mitosis or hard-negative pixels are excluded from the pseudo-mask term. If they are not, the semi-supervised loss actively penalizes correct mitosis predictions—directly suppressing the detection and classification targets. Because the Track 2 classifier consumes segmentation encoder features, this bias can propagate to the classification head. The preliminary-test scores alone cannot discriminate between a model helped by pseudo-masks and one harmed by them.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a teacher–student segmentation framework for the MIDOG 2025 challenge, targeting mitosis detection (Track 1) and atypical mitosis classification (Track 2). The student UNet is trained with online pseudo-masks from a frozen teacher; for Track 1 the teacher is pretrained on PanNuke, while for Track 2 pseudo-masks are derived from classical morphological operations on the H channel after Macenko normalization. Domain generalization is pursued via contrastive learning and a domain-adversarial loss, and a multi-scale CNN classifier head consumes encoder features for Track 2. On the challenge's preliminary test set, the authors report F1 = 0.7660 for Track 1 and balanced accuracy = 0.8418 (abstract: 0.8414) for Track 2, interpreting this as evidence for the effectiveness of the unified framework.","tokens_in":7252,"tokens_out":4697,"duration_ms":49181,"significance":"If the reported scores are reliable, the proposed system is competitive on an externally scored preliminary test set and addresses a practically important problem, namely domain shift in mitosis detection and classification. The paper draws on multiple public datasets and follows the challenge protocol, which aids reproducibility. Its main strengths are the external evaluation, the combination of detection and classification in one framework, and the explicit use of semi-supervised pseudo-masks to mitigate label scarcity. However, the causal claims in the abstract and conclusion go beyond the experiments because no ablation or baseline is reported, and the pseudo-mask generation for Track 2 has a plausible failure mode that is not analyzed.","major_comments":[{"comment":"No ablation or baseline is reported. Tables 3 and 4 report only the final system, so the abstract's claim that the results 'demonstrat[e] the effectiveness of integrating segmentation-based detection and classification into a unified framework' is unsupported. The Discussion itself concedes 'this study is limited by the absence of a detailed ablation analysis to quantify the contribution of each component.' To support the headline claim, the paper needs at least a plain-UNet baseline and one-factor ablations for the pseudo-mask supervision, the contrastive/DANN losses, and the multi-scale classifier head.","section":"Results / Discussion"},{"comment":"For Track 2, pseudo-masks are derived from the H channel of Macenko-normalized images via Gaussian blur, adaptive Otsu thresholding, and morphological opening. The loss description says 'annotated mitosis and hard negatives were combined with teacher-generated pseudo masks for normal nuclei,' but it does not state that annotated mitotic figures are excluded from the pseudo-mask. Since mitotic figures are hematoxylin-dense, they are likely to be included in the thresholded normal-nucleus mask. If so, Lsemi would penalize correct mitosis predictions and could suppress the detection and classification targets. Please specify the mask construction precisely, exclude annotated mitosis/hard-negative pixels from the pseudo-mask if they are not already excluded, or provide empirical evidence that this contamination does not occur.","section":"Frozen Teacher Module / Loss Functions"},{"comment":"The dataset counts are internally inconsistent. Summing the rows in Table 2 gives 3,720 + 11,939 + 3,012 = 18,671 total, not 15,668; the normal column sums to 13,458, not 12,137; the atypical column sums to 3,958, not 3,551. The text says 1,255 Octopath images were excluded, but Table 2 still lists Octopath as 3,012 total. Either the table or the text is wrong, and the total counts do not add up. This must be corrected because the dataset composition is central to reproducing the Track 2 results.","section":"Table 2 / Dataset"},{"comment":"The preliminary test set consists of only 20 cases (four domains, five cases per domain), and the paper reports only point estimates. No confidence intervals, per-case scores, or statistical comparison are given. The per-domain F1 values in Table 3 range from 0.7319 to 0.8649, so the overall difference between in-domain (0.7896) and preliminary (0.7660) performance may be within sampling variability. Please report per-case results or bootstrap confidence intervals, or explicitly acknowledge this limitation in the interpretation of the preliminary scores.","section":"Results / Tables 3 and 4"}],"minor_comments":[{"comment":"The abstract states balanced accuracy 0.8414 for Track 2, while Table 4 and the Results text report 0.8418. Please use one consistent value.","section":"Abstract / Results"},{"comment":"The phrase 'UNet segmentation backbone integrates domain generalization modules' is missing 'that' or 'which'; it should read 'backbone that integrates'.","section":"Introduction / Materials"},{"comment":"There is a subject-verb agreement error: 'can effectively improves' should be 'can effectively improve'.","section":"Discussion"},{"comment":"The labels 'Enconder T', 'Deconder T', and 'Shared netwrk' contain typos; they should be 'Encoder T', 'Decoder T', and 'Shared network'.","section":"Figure 1"},{"comment":"The header 'MIDOG Pre. (overall)' is unclear. Please expand to 'Preliminary (overall)' and clarify that the 'MIDOG' column is the in-domain test.","section":"Table 3"},{"comment":"The sentence '5 cases per each domain' should be '5 cases per domain'. Also, the relationship between the filtered Octopath count and Table 2 should be clarified.","section":"Materials and Methods / Track 2"}],"recommendation":"major_revision","confidential_remarks":"This is a challenge-report style paper. The external preliminary scores are valuable, but the central 'effectiveness' claim requires ablations or at least a more cautious framing. The pseudo-mask contamination issue is a correctness risk that needs either a fix or an empirical demonstration that it does not occur. The dataset-table inconsistencies must be corrected before the paper can be considered publication-ready."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First, this is a standard challenge-system paper: known components (UNet, teacher-student, contrastive learning, DANN, multi-scale classifier) combined sensibly, with a real external preliminary test score. Second, the paper explicitly admits it does no ablation, and the one genuinely novel component—pseudo-masks for normal nuclei—has a design flaw that could make it actively harmful rather than helpful.\n\nWhat's new is the specific configuration and the reported scores on the MIDOG 2025 preliminary test: F1 0.7660 for detection, BA 0.8418 for atypia classification. The authors use multiple public datasets, split at the patient level, and the external prelim set gives some independent check. The architecture is reasonable and the writing is clear. For a challenge report, this is a solid entry.\n\nThe big soft spot, which the stress-test note identifies correctly, is the pseudo-mask supervision for Track 2. The 'teacher' there is not learned; it is classical morphology on the H-channel of Macenko-normalized images—Gaussian blur, adaptive Otsu, morphological opening. Mitotic figures are hematoxylin-dense and will often fall inside the thresholded regions. The paper never states that annotated mitosis or hard-negative pixels are excluded from the pseudo-mask term. If they are not, the semi-supervised loss penalizes correct mitosis predictions, which would directly suppress the detection and classification targets. Even if that is not what happens in practice, the absence of any ablation means we cannot tell whether this term helps or hurts. The paper's own Discussion admits the absence of ablation analysis, so the abstract's phrase 'demonstrating the effectiveness' is too strong.\n\nOther issues are proportionally minor. There are no error bars or case-level analysis for the 20-case prelim set. The Track 2 classification threshold of 0.590 is tuned on the in-domain validation set and then applied to the test, making the balanced-accuracy number an optimistic operating point. Also, the abstract reports BA 0.8414 while the Results table says 0.8418—small, but sloppy.\n\nThe external preliminary scores are still a meaningful signal, but they cannot carry the weight of the effectiveness claim. Who is this for? People entering MIDOG-like challenges or looking for a working baseline in computational pathology. It deserves a referee—it is a legitimate empirical report with a clear limitation statement—but it needs revision before acceptance: at least one ablation with and without pseudo-masks, a plain UNet baseline, and a statement about whether mitotic pixels are masked out of the pseudo-mask loss. My take is conditional accept at best; the scores are plausible, but the claimed effectiveness is not established.","headline":"Competent MIDOG 2025 entry with credible preliminary scores, but no ablation and a questionable morphological pseudo-mask step leave the central effectiveness claim unsupported.","tokens_in":7689,"tokens_out":1773,"would_cite":false,"duration_ms":19117,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single teacher-student UNet detects and classifies mitosis across unseen pathology domains, reporting F1 0.766 and balanced accuracy 0.84 on the MIDOG 2025 preliminary test.","keywords":["mitosis detection","atypical mitosis classification","teacher-student learning","semi-supervised segmentation","domain generalization","histopathology","pseudo-masks","MIDOG 2025"],"falsifier":"Hold out one domain from training and compare F1 with and without the normal-nucleus pseudo-mask supervision; if removing the pseudo-mask branch does not lower F1, the claimed mechanism is not doing the work. Also inspect per-domain false positives on the preliminary test set—especially the domain that dropped to 0.7319 F1—to see whether they are normal nuclei mislabeled as mitosis, which would directly test whether the pseudo-mask supervision reduced the false-positive problem.","tokens_in":6860,"feed_emoji":"🔬","tokens_out":6069,"duration_ms":58516,"temperature":0.7,"pith_summary":"The paper tries to show that mitosis detection is best treated as pixel-level segmentation inside a semi-supervised teacher-student loop, with pseudo-masks for normal nuclei providing dense supervision. It claims this approach copes with scarce mitosis annotations, heavy class imbalance, and the stain/scanner domain shifts that break most histopathology AI tools. For atypical-mitosis classification, the same segmentation encoder feeds a multi-scale CNN head, so detection and classification share one framework. On the preliminary MIDOG 2025 test set the method reports an F1 of 0.7660 for detection and a balanced accuracy of about 0.84 for classification.","feed_headline":"F1 0.766 for cross-domain mitosis detection","feed_subtitle":"A teacher-student UNet with normal-nucleus pseudo-masks also scores 0.84 balanced accuracy on atypical-mitosis classification.","key_machinery":"The load-bearing mechanism is the frozen-teacher-supervised UNet. For Track 1 the teacher is pre-trained on PanNuke to learn nuclear morphology, then synced to the student whenever validation improves; for Track 2 the teacher's normal-nucleus pseudo-masks are hand-crafted from Macenko stain normalization, adaptive Otsu thresholding, and morphological opening. The student sees weakly and strongly augmented views of the same tile; contrastive loss aligns their encoder features, a DANN loss via gradient reversal strips domain-specific signals, and the decoder outputs four classes: background, normal nuclei, mitosis, and hard negatives. A multi-scale CNN classifier with a ResNet-152 refinement b","core_discovery":"On its own terms, the paper claims that a UNet-based teacher-student model with contrastive and domain-adversarial modules can simultaneously detect mitotic figures and classify atypical mitoses across domains. A frozen teacher generates online pseudo-masks—for annotated mitoses, hard negatives, and normal nuclei—while the student learns from weakly and strongly augmented views of the same image. The normal-nucleus pseudo-masks are the key addition: they convert the extreme class imbalance into a dense supervision signal and reduce false positives from non-mitotic nuclei that resemble mitoses. The multi-scale classifier head reuses the encoder features in a multi-task setting for Track 2. Th","pith_inferences":["The paper states that no ablation was run; a direct test of the stated mechanism would be to remove normal-nucleus pseudo-masks, the contrastive loss, or the DANN loss one at a time and measure F1 and balanced accuracy on a held-out domain.","The hand-crafted normal-nucleus pseudo-mask recipe is generic enough that it could transfer to other rare-event detection problems in histology, such as Ki-67-positive cells or tumour-infiltrating lymphocytes, without additional pixel annotations.","The specificity drop from 0.889 in-domain to 0.768 on the preliminary test suggests that per-domain calibration of the classification threshold, rather than a single fixed threshold, could recover some of the lost balanced accuracy.","The uneven per-domain results imply that stain-invariant features only partially close the domain gap; testing on more scanners and staining protocols would reveal whether the residual failures are systematic."],"forward_implications":["A single segmentation-based model can carry both detection and classification, so a deployed mitosis-scoring pipeline needs one network rather than separate detectors and classifiers.","Pseudo-masks for normal nuclei turn class imbalance into dense supervision, which should cut false positives from nuclei that visually mimic mitotic figures.","Contrastive and adversarial losses make the encoder less sensitive to stain and scanner variation, addressing the main known failure mode of AI in histopathology.","The reported degradation pattern—one domain drops hardest and specificity falls more than sensitivity—shows where cross-domain performance is still fragile.","Because the classification head reuses segmentation features, atypical-mitosis classification can be added without training a large standalone classifier, which matters when atypical examples are scarce."],"supporting_citations":[{"why":"Supplies the multi-domain MIDOG++ mitosis bounding-box annotations that Track 1 training and evaluation rely on.","marker":"[8]"},{"why":"Provides PanNuke pixel-level nuclei masks used to pre-train nuclei awareness during the warm-up phase.","marker":"[11]"},{"why":"Defines the UNet encoder-decoder architecture that the whole segmentation framework is built on.","marker":"[17]"},{"why":"Macenko stain normalization whose H channel seeds the classical pseudo-mask generation for normal nuclei in Track 2.","marker":"[19]"},{"why":"Adaptive Otsu thresholding converts the stain-normalized H channel into normal-nucleus pseudo-masks.","marker":"[20]"},{"why":"Contrastive learning loss enforces consistency between weakly and strongly augmented views of the same tile.","marker":"[21]"},{"why":"Domain-adversarial training loss pushes the encoder to discard domain-specific features via gradient reversal.","marker":"[22]"},{"why":"AMi-Br supplies normal-versus-atypical mitosis images used to train the Track 2 classifier.","marker":"[14]"},{"why":"MIDOG 2025 atypical training set provides subtype annotations for MIDOG++ cases in Track 2.","marker":"[15]"},{"why":"Octopath dataset contributes additional atypical and normal mitosis images after filtering out non-mitotic or unknown classes.","marker":"[16]"}],"fun_headline_variants":["Cross-domain mitosis detection F1 0.766","Teacher-student UNet yields F1 0.766 for mitosis","Atypical mitosis classified with 0.84 balanced accuracy","Normal-nucleus pseudo-masks cut mitosis false positives","One model handles mitosis detection and classification"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The training pipeline depends on pseudo-masks for normal nuclei being accurate: for Track 2 these masks are generated by stain normalization plus thresholding and morphological cleanup, and if they systematically mislabel nuclei the student model learns from noisy supervision and the reported cross-domain scores will not reproduce.","fun_headline_variants_meta":{"raw":{"variants":["Cross-domain mitosis detection F1 0.766","Teacher-student UNet yields F1 0.766 for mitosis","Atypical mitosis classified with 0.84 balanced accuracy","Normal-nucleus pseudo-masks cut mitosis false positives","One model handles mitosis detection and classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000347,"raw_usage":{"total_tokens":1767,"prompt_tokens":804,"completion_tokens":963,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":884}},"tokens_in":548,"tokens_out":963,"duration_ms":8965,"temperature":1.0,"reasoning_tokens":884,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:47:33.394927+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out one domain from training and compare F1 with and without the normal-nucleus pseudo-mask supervision; if removing the pseudo-mask branch does not lower F1, the claimed mechanism is not doing the work. Also inspect per-domain false positives on the preliminary test set—especially the domain that dropped to 0.7319 F1—to see whether they are normal nuclei mislabeled as mitosis, which would directly test whether the pseudo-mask supervision reduced the false-positive problem.","supporting_citations":[{"cited_title":"A comprehensive multi-domain dataset for mitotic figure detection","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-domain MIDOG++ mitosis bounding-box annotations that Track 1 training and evaluation rely on."},{"cited_title":"Pannuke: an open pan-cancer histology dataset for nuclei instance segmentation and clas- sification","cited_arxiv_id":null,"evidence_quote":"Provides PanNuke pixel-level nuclei masks used to pre-train nuclei awareness during the warm-up phase."},{"cited_title":"A method for normalizing histology slides for quantitative analysis","cited_arxiv_id":null,"evidence_quote":"Macenko stain normalization whose H channel seeds the classical pseudo-mask generation for normal nuclei in Track 2."},{"cited_title":"A threshold selection method from gray-level histograms","cited_arxiv_id":null,"evidence_quote":"Adaptive Otsu thresholding converts the stain-normalized H channel into normal-nucleus pseudo-masks."},{"cited_title":"A simple framework for contrastive learning of visual representations","cited_arxiv_id":null,"evidence_quote":"Contrastive learning loss enforces consistency between weakly and strongly augmented views of the same tile."},{"cited_title":"Domain-adversarial training of neu- ral networks","cited_arxiv_id":null,"evidence_quote":"Domain-adversarial training loss pushes the encoder to discard domain-specific features via gradient reversal."},{"cited_title":"His- tologic dataset of normal and atypical mitotic figures on human breast cancer (ami-br)","cited_arxiv_id":null,"evidence_quote":"AMi-Br supplies normal-versus-atypical mitosis images used to train the Track 2 classifier."},{"cited_title":"A dataset of atypical vs normal mitoses classification for midog - 2025, April 2025","cited_arxiv_id":null,"evidence_quote":"MIDOG 2025 atypical training set provides subtype annotations for MIDOG++ cases in Track 2."},{"cited_title":"Omg-octo atypical: A refinement of the original omg-octo database to incorporate atypical mitoses, July 2025","cited_arxiv_id":null,"evidence_quote":"Octopath dataset contributes additional atypical and normal mitosis images after filtering out non-mitotic or unknown classes."}],"review_version":1}