{"id":"cd07d571-b271-4c16-892c-7dd051737295","arxiv_id":"2509.02597","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A cascade of FCOS and a ResNet50/EfficientNet ensemble raises detection F1 to 0.8432 and a ConvNeXt ensemble reaches 0.86 balanced accuracy on MIDOG 2025.","lead":"This paper describes the authors' pipeline for the MIDOG 2025 mitosis detection and atypical classification challenge: a two-stage detector-classifier for finding mitotic figures, and an ensemble of standard neural networks for classifying them as atypical or normal. It reports validation F1 and balanced accuracy numbers, but the supporting evidence has internal inconsistencies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4's F1 gain is compromised by an unreported threshold, a matching AP drop, and a baseline that appears to be CMC+MIDOG++, not MIDOG++ as stated.","rationale":"The paper is a straightforward MIDOG 2025 challenge report. The detection contribution is the two-stage pipeline, and its only evidence is Table 4. The table's baseline row duplicates Table 2's CMC+MIDOG++ row, so either the method description or the experiment is misreported. The accompanying AP drop is a red flag: F1 is threshold-dependent and no threshold is given. The atypical-classification section is better supported (4-fold CV with std, preliminary evaluation), but it is not the central claimed contribution. The reader's conditional verdict is appropriate; no new reason to reject outright, but the detection claim needs concrete fixes. I would not accept as-is.","tokens_in":7138,"tokens_out":5798,"duration_ms":60476,"concrete_test":"Re-run the detection experiment with three fixes: (1) train FCOS separately on MIDOG++ only and on CMC+MIDOG++, and apply the same ResNet50/EfficientNetB2 classifier to both detectors; (2) report F1 and AP at several classifier thresholds (e.g., 0.3/0.5/0.7) and full PR curves, using a fixed detector threshold; (3) repeat over at least 5 random 80/20 splits (or use 5-fold CV) and report mean±std. Also verify that no patch used to train the classifier comes from a MIDOG++ case in the detection validation split. If the F1 gain vanishes under any of these checks, or AP systematically decreases, Table 4's claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central detection claim—that the two-stage detection–classification framework improves robustness (F1 0.803→0.8432, Table 4)—rests on a comparison with internal inconsistencies. The baseline row in Table 4 (F1=0.803, AP=0.842) exactly duplicates Table 2's 'CMC + MIDOG++' row, but the text says the baseline was trained on MIDOG++; Table 2's MIDOG++-only row is F1=0.794, AP=0.782. Thus the reported F1 gain may conflate the cascade's effect with the effect of adding CMC training data. Independently, the cascade lowers AP from 0.842 to 0.787. F1 is computed at a single operating threshold, and no threshold is reported; a re-ranking cascade can improve F1 at one threshold while degrading the precision-recall ranking (AP). Without a threshold, PR curves, or a matched-recall comparison, the F1 increase does not demonstrate 'enhanced robustness.' The single 80/20 split, with no cross-validation or error bars, further limits the claim. Credit is due for the atypical-classification section, which reports cross-validation with standard deviations, but the detection claim specifically is under-supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports the authors' solutions to the two MIDOG 2025 challenge tasks. For mitotic figure detection, they propose a two-stage pipeline: an FCOS detector first proposes candidate regions, and a secondary classifier (ResNet-50 or an ensemble of ResNet-50 and EfficientNet-B2) re-evaluates each candidate to reduce false positives. They report that this cascade improves F1 from 0.803 to 0.8432 on a validation split of MIDOG++. For atypical mitosis classification, they compare several CNN and ViT backbones using 4-fold cross-validation and then ensemble the ConvNeXt family, reporting a balanced accuracy of 0.8794 in cross-validation and 0.86 on the official preliminary evaluation. The paper is written as a challenge-participation report, with implementation details and public datasets.","tokens_in":7404,"tokens_out":3603,"duration_ms":43467,"significance":"If the detection result were properly supported, the main contribution would be a practical demonstration that a detection-classification cascade can boost mitotic-figure detection F1 beyond a strong one-stage detector, which is of interest to the computational pathology community. The atypical classification part is more carefully evaluated: it uses 4-fold cross-validation with standard deviations, compares several modern architectures, and reports an official preliminary evaluation score. The paper also contributes useful details about training data combinations (MIDOG++, CMC, CMCCT) and an honest limitation that no external validation was used during development. However, the central detection claim is currently undermined by an internal inconsistency in the baseline comparison and by the absence of operating-point and variability analyses.","major_comments":[{"comment":"The baseline row in Table 4 (F1=0.803, AP=0.842) is identical to the 'CMC + MIDOG++' row in Table 2, while the text states the baseline was 'trained on MIDOG++'. Table 2 reports MIDOG++-only as F1=0.794, AP=0.782. If the baseline actually used CMC+MIDOG++, then the F1 gain from the cascade (0.803 to 0.8432) is confounded with the gain from adding CMC training data; if the baseline was indeed MIDOG++-only, the table numbers are inconsistent. This must be corrected and the comparison re-run before the central detection claim can be evaluated.","section":"Table 4 and 'Validation of detection-classification framework'"},{"comment":"The cascade increases F1 from 0.803 to 0.8432 but decreases AP from 0.842 to 0.787. F1 is computed at a single decision threshold, and the classifier threshold is not reported. A re-ranking cascade can improve F1 at one threshold while degrading the precision-recall ranking, as AP indicates here. Without reporting the threshold, precision-recall curves, or a matched-recall comparison, the F1 improvement alone does not demonstrate 'enhanced detection robustness'. Please provide operating-point details and a threshold-independent analysis (e.g., PR curves or AP with confidence intervals).","section":"Table 4, F1 vs AP"},{"comment":"All detection results are based on one fixed 80/20 split, with no cross-validation, repeated runs, or error bars. The classification patch split (7:1:2) in Table 3 also appears to be a single split. Given the small F1 differences involved (e.g., 0.794 vs 0.803) and the threshold sensitivity noted above, the manuscript should report variability across at least a few seeds or split folds, or explicitly state that these are single-run values and avoid making strong comparative claims from them.","section":"Tables 1-4, single split"},{"comment":"The classifier in the cascade is trained on 50x50 patches extracted from MIDOG++ with a 7:1:2 split, while the detector is validated on an 80/20 split of MIDOG++. The paper does not state whether the classifier training/validation patches were separated from the detector validation split. If the classifier saw patches from images that also appear in the detector's validation set, this would leak information and inflate the reported F1. Please clarify the relationship between the patch split and the detection split, and if necessary re-evaluate with strictly disjoint splits.","section":"'Validation of different classification models' and Table 4"},{"comment":"The final atypical-classification ensemble is described only as 'the family of ConvNeXt models'. It is not specified which models (ConvNeXt_large, ConvNeXt_base, ConvNeXt_base+CBAM?) are included, how many folds contributed, or how the ensemble was selected. Moreover, the cross-validation table used for model selection is also presented as the main result, which risks selection bias. The official preliminary evaluation score (0.86) is reported without confidence intervals or details about the evaluation set size. Please specify the ensemble construction and report the official evaluation with as much metadata as available.","section":"'Preliminary evaluation' and 'Model ensemble'"}],"minor_comments":[{"comment":"The sentence ending 'similarity to impostor cells (e.g., apoptotic cells or lymphocytes).' is grammatically incomplete; a verb or clause is missing.","section":"Introduction"},{"comment":"Table 3 reports no standard deviations or confidence intervals, whereas Table 5 does. Adding variability estimates to Table 3 would help interpretation.","section":"Table 3"},{"comment":"EfficientNetB1's balanced accuracy is reported as 0.8325 ± 0.0000, which suggests identical splits or a rounding artifact; please verify.","section":"Table 5"},{"comment":"Several references contain formatting artifacts, e.g., 'Y equan Bie' instead of 'Yequan Bie'.","section":"References"},{"comment":"The text says 'one NVIDIA GeForce RTX 4090 GPUs' but singular/plural agreement is inconsistent.","section":"Implementation details"}],"recommendation":"major_revision","confidential_remarks":"This is a challenge-solution report that could be acceptable for a workshop or short paper venue after revision. The main barrier is the internal inconsistency between Table 2 and Table 4, which directly affects the paper's headline detection claim. The authors should be asked to clarify the baseline, provide threshold-independent evaluation, and add variability estimates. I do not see a need for rejection, as the issues appear fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a MIDOG 2025 challenge write-up. The detection half doesn't support its own headline. The atypical classification half is more careful.\n\nWhat's actually useful: the paper gives a clean comparison of a bunch of backbones for atypical mitosis classification (Table 5, with mean±std over 4-fold CV), and the ConvNeXt result at 0.8794 balanced accuracy is a reasonable data point. The ensemble of ConvNeXt models reaching 0.86 on the official preliminary set is a genuine challenge result. The detection pipeline is a standard two-stage cascade (FCOS + ResNet50 classifier), which is basically ref 10 applied to MIDOG. No new architecture, no public code or data, so novelty is thin.\n\nThe soft spot is the main detection claim. The text says the Table 4 baseline was FCOS trained on MIDOG++, with F1=0.803 and AP=0.842. But Table 2 shows that exact F1/AP pair belongs to CMC+MIDOG++ training, while the MIDOG++-only row is F1=0.794, AP=0.782. So the reported improvement from 0.803 to 0.8432 conflates adding the cascade with adding CMC training data. That is a load-bearing inconsistency, not a footnote. Also, the cascade decreases AP from 0.842 to 0.787—meaning a threshold-specific F1 gain while ranking quality degrades. Since no decision threshold is reported and there are no PR curves or error bars (single 80/20 split), the 'enhanced robustness' claim is not established. The atypical classification section, by contrast, uses cross-validation with standard deviations, which is the right level of rigor for this kind of report.\n\nWho's this for? People following the MIDOG challenge and wanting a quick sense of what worked on the atypical classification track. The detection section should be read with caution until the authors fix the training-set description and publish thresholds or PR analysis.\n\nRecommendation: worth sending to peer review, because the atypical track contribution is legitimate and the detection inconsistency is fixable, but it needs a revision that addresses the baseline mismatch and the AP drop.","headline":"A plausible challenge report undone by an inconsistent baseline in its main detection claim; the atypical classification section is more solid.","tokens_in":7903,"tokens_out":2006,"would_cite":false,"duration_ms":20628,"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":"Two-stage cascade detects mitotic figures at F1 0.8432","keywords":["mitotic figure detection","atypical mitosis classification","computational pathology","two-stage detection","model ensemble","FCOS","ConvNeXt","whole-slide histology"],"falsifier":"Run the same detection pipeline under repeated cross-validation on the MIDOG++ dataset, reporting F1 and AP as distributions, and plot the precision-recall curves of the detector alone versus detector-plus-classifier at matched recall. If the cascade's F1 advantage does not persist across folds or shifts with threshold, the claimed benefit is not robust.","tokens_in":6998,"feed_emoji":"🔬","tokens_out":4685,"duration_ms":52374,"temperature":0.7,"pith_summary":"The paper argues that mitotic figure detection in cancer pathology improves when a fast, sensitive detector proposes candidates and a separate image classifier then filters them, removing lookalike cells that confuse single-stage models. On a validation split of the challenge's multi-domain dataset, adding a ResNet50-and-EfficientNet ensemble classifier raised F1 from 0.803 to 0.8432, at the cost of lower average precision. For the companion task of distinguishing atypical from normal mitotic figures, the paper shows that averaging predictions across several ConvNeXt-family models reaches 0.8794 balanced accuracy in cross-validation. The work is a challenge solution: its claims are measured on defined training and validation splits, not a new biological mechanism.","feed_headline":"Cascade classifier lifts mitotic figure F1 to 0.8432","feed_subtitle":"A patch-level second model filters lookalike cells that fool single-stage detectors; ensembled models also classify atypical mitoses.","key_machinery":"The two-stage detection-classification cascade: stage one is the FCOS anchor-free detector predicting centers, box offsets, and objectness; stage two crops each positive candidate to a patch and runs a ResNet50/EfficientNet ensemble classifier that keeps only confirmed candidates. The companion machinery is probability averaging across independently trained models for atypical/normal classification, with the ConvNeXt+CBAM family as the strongest single backbones.","core_discovery":"The central claim is that detection and classification should be separated: an anchor-free FCOS detector is tuned for high recall, and a dedicated convolutional classifier re-examines each proposed patch to suppress impostor cells such as apoptotic bodies and lymphocytes. The paper reports that this cascade raises F1 from 0.803 to 0.8432 on its validation split while noting AP decreases from 0.842 to 0.787, and that ensembling two classifiers outperforms either alone. For atypical classification, the paper finds convolutional models, especially the ConvNeXt family with attention, outperform a Vision Transformer on limited data, and that ensembling the top convolutional models gives the best","pith_inferences":["Editorially, the F1 improvement may partly reflect threshold selection: AP falls from 0.842 to 0.787 after the cascade, so the F1 gain could shrink or reverse at matched operating points; a precision-recall curve comparison would settle it.","Editorially, all detection comparisons rest on one 80/20 split; repeated splits or cross-validation would be needed to confirm that the cascade gain is not split-specific.","Editorially, the same cascade idea could be tested on other small-object detection tasks in whole-slide histology, such as detecting other rare cell events, where a recall-first detector plus a discriminative patch classifier is a natural fit."],"forward_implications":["If the cascade generalizes, single-stage histology detectors can be improved by a separate patch-level filter without retraining the detector.","The F1 gain indicates false positives from morphologically similar non-mitotic cells are the bottleneck, and a classifier trained on curated positives and negatives directly attacks that bottleneck.","The finding that adding canine breast-cancer data helps but canine mast-cell-tumor data hurts suggests domain composition of training data is a first-order choice.","Ensembling convolutional backbones provides more reliable atypical/normal scores than any single network, which matters for clinical grading where false negatives are costly."],"supporting_citations":[{"why":"Supplies the FCOS anchor-free detector used as stage one of the cascade.","marker":"(15)"},{"why":"Supplies the ResNet backbones used as the candidate re-classifier.","marker":"(16)"},{"why":"Provides the MIDOG++ multi-domain dataset on which detection validation is performed.","marker":"(12)"},{"why":"Canine mast-cell tumor dataset whose addition to training hurt detection performance.","marker":"(13)"},{"why":"Canine breast cancer dataset whose addition to training improved detection.","marker":"(14)"},{"why":"Primary training set for atypical mitotic figure classification.","marker":"(19)"},{"why":"Supplementary atypical/normal histology dataset used to augment atypical classification training.","marker":"(20)"},{"why":"ConvNeXt architectures that give the strongest single-model results for atypical classification.","marker":"(25)"},{"why":"CBAM attention module that improves ConvNeXt_base in atypical classification.","marker":"(26)"},{"why":"EfficientNet supplies the second ensemble classifier and the scaling-family ablation results.","marker":"(27)"}],"fun_headline_variants":["Two-stage cascade lifts mitotic F1 to 0.8432","Mitotic F1 up, AP down: trade-off of cascade approach","Ensembled convnets outperform ViT for atypical mitoses","Split detection and classification cuts lookalike false positives"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The detection results rely on one fixed training/validation split with no error bars, so the F1 boost from the added classifier could be an artifact of that split or the score threshold rather than a stable gain.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage cascade lifts mitotic F1 to 0.8432","Mitotic F1 up, AP down: trade-off of cascade approach","Ensembled convnets outperform ViT for atypical mitoses","Split detection and classification cuts lookalike false positives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000803,"raw_usage":{"total_tokens":3314,"prompt_tokens":643,"completion_tokens":2671,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":387,"completion_tokens_details":{"reasoning_tokens":2598}},"tokens_in":387,"tokens_out":2671,"duration_ms":24172,"temperature":1.0,"reasoning_tokens":2598,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:00:26.764995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same detection pipeline under repeated cross-validation on the MIDOG++ dataset, reporting F1 and AP as distributions, and plot the precision-recall curves of the detector alone versus detector-plus-classifier at matched recall. If the cascade's F1 advantage does not persist across folds or shifts with threshold, the claimed benefit is not robust.","supporting_citations":[],"review_version":1}