{"id":"f7ef8095-4fa1-4f1d-a71d-6ae8328adb33","arxiv_id":"2607.28007","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Frozen pathology FMs (especially H-optimus-0 and Virchow) support mitotic-figure detection competitively with end-to-end ResNet50 and transfer slightly better to TUPAC16.","lead":"Frozen pathology foundation models can act as detection backbones for finding mitotic figures, nearly matching a fully trained ResNet50 and slightly beating it out-of-domain. This matters because it shows image-level self-supervised features already carry the spatial detail needed for dense cell detection without retraining the encoder.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified; the hedged claim is supported by the frozen-ResNet control and overlapping-CI results.","rationale":"The reader correctly flags the single-scale ViT → ViTDet neck as the least secure methodological assumption and correctly notes soft rankings under bootstrap overlap plus a single OOD set. Those points limit how strongly one can rank individual backbones or assert OOD superiority, but they do not undercut the paper’s carefully hedged central claim (suitable / competitive / may be slightly more robust). The frozen-ResNet-50 control and multi-head design already address the main alternative explanation (any frozen stem + learnable neck would suffice). Public code, patient-level splits, and transparent limitations further keep correctness risk low. I therefore leave the verdict at ACCEPT; no adjustment is warranted. Agreement with the reader is partial because their weakest_assumption is real but not decisive against the claim as written.","tokens_in":13091,"tokens_out":560,"duration_ms":32723,"concrete_test":"Re-run the top three frozen FM configs (H-optimus-0/Virchow/Virchow2 + RetinaNet) with a higher-capacity neck ablation (e.g., deeper ViTDet FPN or simple bilinear multi-scale pyramid with 2× channel width) under identical seeds; if best frozen F1 rises by >0.02 in-domain and the OOD edge over fine-tuned ResNet-50 disappears or reverses, the ‘representations already sufficient’ reading weakens and neck capacity becomes the real bottleneck.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader’s neck concern is the softest design point, but it is not load-bearing against the paper’s actual claim. The claim is only that frozen FM latents are already ‘suitable’ / competitive for dense MF detection (not that the neck is optimal or that FMs dominate). A frozen ResNet-50 under the same neck-and-head-only regime trails the strongest frozen FMs both in-domain and OOD (Tables 1–2), which isolates pathology pretraining from the mere act of freezing. In-domain the fine-tuned ResNet-50+RetinaNet still leads on point estimate (F1 0.7917 vs 0.7718); OOD the ranking flips on point estimate (0.7349 vs 0.7193) with substantial CI overlap. The discussion already flags frozen-only evaluation, single secondary OOD set, and marginal CIs. No internal inconsistency or unsupported leap is required for the qualified conclusion to hold.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The manuscript asks whether frozen latent spaces of current pathology foundation models (UNI, UNI2-h, Virchow, Virchow2, H-optimus-0, H-optimus-1), all pretrained with image-level self-supervision, already encode features that are discriminant and spatially resolved enough for dense mitotic-figure (MF) detection. Six frozen ViT FMs are paired with three detector heads (Faster R-CNN, RetinaNet, Deformable DETR) via a lightweight ViTDet-style feature-pyramid neck and compared to ImageNet ResNet-50 baselines in both fully fine-tuned and frozen regimes on MIDOG++ (in-domain) and TUPAC16 (OOD). On MIDOG++, the best frozen configuration (H-optimus-0 + RetinaNet, F1 0.7718) approaches but does not surpass fine-tuned ResNet-50 + RetinaNet (F1 0.7917); on TUPAC16 the ranking of point estimates reverses (0.7349 vs 0.7193). A frozen ResNet-50 control lags the strongest FMs in both settings. The authors conclude that FM latents are suitable for direct MF detection and may transfer slightly more robustly under domain shift. Code is released.","tokens_in":13306,"tokens_out":1323,"duration_ms":33299,"significance":"The work cleanly addresses an open and practically relevant question: whether pathology FMs, already strong on classification, can serve as frozen detection encoders for a clinically used dense task. Strengths include a controlled bake-off design (three detector paradigms, fine-tuned and frozen ResNet-50 controls, patient-level splits, bootstrap CIs, micro-F1/precision/recall plus FROC), multi-domain and secondary OOD evaluation, and public code. The frozen-ResNet control usefully isolates pathology-specific pretraining from the mere act of freezing. If the qualified claim holds, it supports using frozen FMs as drop-in detection stems and motivates parameter-efficient adaptation studies. The contribution is empirical rather than methodological, but the evidence is of direct interest to computational pathology and foundation-model evaluation.","major_comments":[{"comment":"Tables 1–2 and §3–4: the OOD claim that frozen FMs are “slightly more robust” rests on point-estimate F1 reversal (H-optimus-0+RetinaNet 0.7349 vs ResNet-50+RetinaNet 0.7193) with substantial bootstrap CI overlap. The discussion already notes marginal per-configuration CIs and a single secondary OOD set, but the abstract and conclusion still state the robustness advantage without that qualification. Please either (i) add a paired/slide-level significance test or explicit non-inferiority framing, or (ii) soften the abstract/conclusion wording to match the hedged discussion so the central transfer claim is not overstated.","section":"Abstract; §3 Results; §4 Discussion; Tables 1–2"},{"comment":"§2.2–2.3: the design discards non-spatial tokens and relies on a single learnable ViTDet-style neck to build the multi-scale pyramid from a frozen single-scale token map. The frozen-ResNet control (same neck/head-only regime) trailing the best FMs supports attributing gains to pathology pretraining, so this is not fatal to the “suitable” claim. Nonetheless, neck capacity remains a soft confounder for any residual gap to the fine-tuned baseline. A brief ablation (e.g., neck depth/channel width, or reporting neck parameter count vs head) or an explicit limitation sentence that gaps may partly reflect neck under-capacity would make the isolation argument tighter without expanding scope.","section":"§2.2–2.3; §4 Discussion"}],"minor_comments":[{"comment":"Abstract and intro cite the public GitHub URL; the title-page footnote still points at an anonymous.4open link. Align the camera-ready link.","section":"Title page / Abstract"},{"comment":"§2.1: “1008 and 1024px for patch-14 and patch-16” is clear, but state explicitly which backbones use which patch size (and whether input was center-cropped or resized) so the stride tables in §2.3 are reproducible without the code.","section":"§2.1"},{"comment":"Table 1 vs Table 2: FROC is reported on a [0,8] FP/image scale and can exceed 1; a one-line reminder in the table caption that FROC-AUC here is not bounded by 1 would avoid reader confusion.","section":"Tables 1–2"},{"comment":"§2.5 Evaluation: “free-range operating characteristic” appears to be a typo for “free-response operating characteristic” (FROC).","section":"§2.5"},{"comment":"Related work briefly notes CellViT++ underperforming a from-scratch RetinaNet on MIDOG++; a short clause on how the present frozen-backbone+detection-head setup differs (no instance decoder, MF-specific heads, different tiling) would help readers place the result.","section":"§1 Introduction"},{"comment":"Minor typography: “F rozen”, “F eature”, “T raining” section headings show stray spaces; “Severallimitationstempertheseconclusions” and similar run-ons in §4 need spacing fixes.","section":"§2–4"}],"recommendation":"minor_revision","confidential_remarks":"Solid, reproducible empirical paper; I agree with the reader that the neck concern is real but not load-bearing against the hedged claim. Minor revision is mainly to align abstract/conclusion hedging with the CI-overlapping OOD evidence. Suitable for a methods/applications track in medical image analysis or computational pathology."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this is a careful bake-off, not a claim of dominance. Six frozen pathology FMs (UNI family, Virchow family, H-optimus) plus three detector heads beat or match a frozen ResNet50 and come close to a fully fine-tuned ResNet50+RetinaNet on MIDOG++; on TUPAC16 the best frozen setup (H-optimus-0 + RetinaNet) edges the fine-tuned baseline on point-estimate F1. That is new relative to the classification and CellViT-style work they cite, and they isolate the pathology pretraining with the frozen-ResNet control.\n\nWhat they do well: patient-level splits, shared protocol, three detector families so the result is not head-specific, bootstrap CIs, micro-F1 plus FROC, public code, and a discussion that already flags the real limits (frozen-only, one secondary OOD set, overlapping intervals). The central claim is hedged exactly to the evidence—“suitable / competitive / slightly more robust”—not “FMs win.” Rankings are mostly stable across heads and datasets; H-optimus-0 and Virchow lead the frozen pack.\n\nSoft spots are real but proportionate. The ViTDet-style neck is doing a lot of work turning single-scale tokens into a pyramid; if the neck is under-powered the “latent space is already suitable” reading is a bit softer than it looks. They do not try LoRA or any backbone adaptation, so we do not know the ceiling. OOD is one dataset with substantial CI overlap, so the robustness story is suggestive, not settled. None of that breaks the result.\n\nThis is for people building detection pipelines on pathology FMs or running mitosis challenges. Methods are transparent enough that a serious referee should see it. I would bring it to reading group if we are talking FM transfer or MIDOG-style benchmarks, cite it when I need the frozen-backbone detection numbers, and send it to review rather than desk-reject. Solid applied evidence; engage with it.","headline":"Clean empirical answer: frozen pathology FM latents are already good enough for dense mitotic-figure detection and look slightly more robust OOD than a fine-tuned ResNet50.","tokens_in":14029,"tokens_out":526,"would_cite":true,"duration_ms":17194,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Frozen pathology foundation models can detect mitotic figures almost as well as a fully trained detector, and hold up slightly better out of domain.","keywords":["pathology foundation models","mitotic figure detection","object detection","frozen backbones","MIDOG++","domain generalization","RetinaNet","ViTDet neck"],"falsifier":"Fine-tune the same foundation backbones with a light adapter (for example LoRA) under identical necks and heads; if the in-domain gap to the ResNet-50 baseline closes substantially while out-of-domain robustness collapses, the claim that the frozen spaces are already sufficient is weakened.","tokens_in":13922,"feed_emoji":"🔬","tokens_out":890,"duration_ms":16485,"temperature":0.7,"pith_summary":"Mitotic figures—cells caught in the act of dividing—are counted by pathologists to grade many cancers, but the work is slow and inconsistent. Modern pathology foundation models already work well when you ask them to classify a small patch as mitotic or not. This paper asks the harder question: can their frozen internal features also localize those cells in a full image, the way a dedicated object detector does? The authors plug six frozen foundation-model backbones into three standard detector heads and compare them with a ResNet-50 baseline trained end-to-end on the multi-domain MIDOG++ set, then re-test on the TUPAC16 breast-cancer set as an out-of-domain check. The best frozen models come close to the fully trained baseline in-domain and slightly surpass it out-of-domain, showing that image-level self-supervision already leaves enough spatially resolved signal for dense mitotic-figure detection.","feed_headline":"Frozen pathology models nearly match trained mitotic detectors","feed_subtitle":"Image-level self-supervision already leaves enough spatial signal to localize dividing cells, and holds up better out of domain","key_machinery":"Frozen ViT foundation-model backbones whose non-spatial tokens are discarded and whose single-scale patch-token maps are lifted into a multi-scale pyramid by a lightweight ViTDet-style neck, then fed to standard detector heads (RetinaNet, Faster R-CNN, Deformable DETR).","core_discovery":"The frozen latent spaces of current pathology foundation models, all trained only with image-level self-supervision, already contain features that are both discriminant and spatially resolved enough to drive mitotic-figure object detection competitively with an end-to-end-trained ResNet-50 baseline, and they transfer slightly more robustly under domain shift.","pith_inferences":["If the neck is the real bottleneck, stronger multi-scale adapters or hierarchical ViT pretraining could push frozen FMs past fully trained convolutional detectors without unfreezing the backbone.","The slight OOD edge suggests frozen FMs may be preferable whenever the target organ or staining protocol is poorly represented in the labeled detection set.","Parameter-efficient fine-tuning is the natural next experiment; success would turn the residual in-domain gap into a solvable engineering problem rather than a representation limit."],"forward_implications":["Mitotic-figure detectors can be built by freezing a pathology foundation model and training only a neck and head, cutting labeled-data and compute cost relative to full end-to-end training.","Among current public models, H-optimus-0 and Virchow are the strongest ready-made detection encoders for this task.","Out-of-domain mitotic counting may benefit more from frozen large-scale features than from domain-adapted convolutional stems.","The same frozen-backbone recipe can be tried on other dense pathology tasks that currently rely on fully fine-tuned ImageNet stems."],"fun_headline_variants":["Frozen pathology FMs rival end-to-end mitotic detectors","Image-level FM latents suffice for mitotic figure detection","Pathology foundation models double as mitotic detectors","Frozen FM features drive competitive mitosis detection","Self-supervised pathology FMs hold up better out of domain"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That a small learned neck alone can turn a frozen single-scale patch grid into the multi-scale features modern detectors need, so any remaining performance gap truly reflects the pretrained features rather than the neck’s limits.","fun_headline_variants_meta":{"raw":{"variants":["Frozen pathology FMs rival end-to-end mitotic detectors","Image-level FM latents suffice for mitotic figure detection","Pathology foundation models double as mitotic detectors","Frozen FM features drive competitive mitosis detection","Self-supervised pathology FMs hold up better out of domain"]},"model":"grok-4.5","effort":"low","cost_usd":0.004486,"raw_usage":{"total_tokens":1328,"prompt_tokens":816,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":44864000,"prompt_tokens_details":{"text_tokens":816,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":454,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":816,"tokens_out":58,"duration_ms":8105,"temperature":1.0,"reasoning_tokens":454,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T20:35:46.446426+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Fine-tune the same foundation backbones with a light adapter (for example LoRA) under identical necks and heads; if the in-domain gap to the ResNet-50 baseline closes substantially while out-of-domain robustness collapses, the claim that the frozen spaces are already sufficient is weakened.","supporting_citations":[],"review_version":1}