{"id":"fab48bf5-9824-4718-b1e6-79cf6114f506","arxiv_id":"2502.00408","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PathoSAM, a finetuned Segment Anything Model trained on six histopathology datasets, achieves state-of-the-art automatic and interactive nucleus instance segmentation across diverse tissue types.","lead":"The authors trained a version of the Segment Anything Model, called PathoSAM, on six histopathology datasets so that it segments cell nuclei accurately in tissue images. It works both automatically and interactively with user clicks, and the authors report that it beats existing methods across many tissue types and staining conditions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OOD evidence may be contaminated: Lizard, a training dataset, is composed in part of CoNSeP images that are also used as an out-of-domain evaluation set, so the 'unseen dataset' generalization claim may rely on leaked images.","rationale":"I read the paper as a comprehensive empirical adaptation of SAM to histopathology. The open-source code, tool integration, and multi-dataset evaluation are real strengths. The reader's weakest assumption, annotation-protocol consistency, is plausible, but I see a more load-bearing issue: whether the out-of-domain evaluation is actually out-of-domain. Lizard is commonly described as a merger of existing colorectal datasets, including CoNSeP, and CoNSeP is labeled as an out-of-domain evaluation dataset in Section 2.5. If Lizard's training split contains images from CoNSeP, then the 'unseen dataset' claim is not valid for CoNSeP, and the reported OOD generalization results for that dataset could be inflated. This is directly checkable from public dataset metadata and the released training splits. The lack of error bars and the tuning of AIS thresholds on a single OOD validation set remain secondary but do not change my overall recommendation: the paper should be accepted only after the overlap check is performed and the reported numbers are corrected or explicitly justified. I therefore keep the reader's CONDITIONAL verdict, with an additional condition on image-level de-duplication.","tokens_in":15797,"tokens_out":13322,"duration_ms":132892,"concrete_test":"Obtain the exact Lizard training split used for the PathoSAM generalists and the CoNSeP evaluation images (App. D.1/D.2). Compute image-level overlap by matching original filenames, slide IDs, or perceptual hashes of 512x512 tiles with the same extraction coordinates. Report the number and percentage of overlapping evaluation images. Then re-run the automatic and interactive evaluations with all overlapping images removed, using the same watershed thresholds, and compare the per-dataset and averaged mean segmentation accuracies. If the overlap is zero, the concern does not land; if it is nonzero and removing the overlapping images changes the winner on CoNSeP or lowers the OOD average, the 'unseen generalization' and SOTA claims must be re-stated with the de-contaminated numbers.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that a single PathoSAM model generalizes to unseen histopathology datasets and is state-of-the-art for automatic and interactive nucleus segmentation. Section 2.5 defines six H&E training datasets, including Lizard (Graham et al., 2021), and separately lists CoNSeP (Graham et al., 2019) as one of the 'remaining 8 datasets' used for out-of-domain evaluation, i.e. data 'not directly represented in the training set'. Lizard is a merged, re-annotated collection built from existing colorectal datasets, and its constituent image sources include CoNSeP (as well as CPM-15, CPM-17 and PanNuke). If the Lizard training split contains any of the same slides or image tiles that later appear in the CoNSeP evaluation, then CoNSeP is not out-of-domain at all. The same possibility applies to MoNuSAC relative to MoNuSeg. The manuscript does not report an image-level de-duplication check. Because the abstract's strongest claim is 'new state-of-the-art' and the evidence for it includes 'remarkable generalization' on out-of-domain data, a confirmed overlap would remove one of the main independent pillars of the claim; the in-domain comparisons alone may not establish a state-of-the-art model for unseen data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PathoSAM, a finetuned variant of the Segment Anything Model (SAM) for nucleus segmentation in histopathology. It supports interactive segmentation from point/box prompts and automatic instance segmentation via an additional decoder with seeded watershed post-processing (AIS), following the microSAM recipe. The authors train generalist ViT-B/L/H models on a combination of six H&E datasets and evaluate on 12 datasets (6 in-domain test splits and 6 out-of-domain), comparing against HoVerNet, StarDist, CellViT, HoVerNeXt, InstanSeg, SAM, and microSAM. They report that PathoSAM is the new state-of-the-art for automatic and interactive nucleus instance segmentation, while for semantic segmentation on PanNuke it ranks second to CellViT. The paper also presents a user study with QuPath and napari integration, whole-slide-image inference scripts, and open-source code and model weights.","tokens_in":16085,"tokens_out":10615,"duration_ms":102202,"significance":"If the results hold, PathoSAM is a practically valuable contribution: a single open-source model that performs automatic and interactive nucleus segmentation across diverse histopathology datasets without per-dataset retraining, with integration into popular annotation tools. The study's strengths include held-out test splits, best-available-version reporting for multi-version baselines, and public code and models. However, the strength of the state-of-the-art claim is currently weakened by the absence of uncertainty estimates or significance tests, and by a potential overlap between the Lizard training data and the CoNSeP out-of-domain evaluation set, which directly bears on the 'unseen dataset' generalization claim.","major_comments":[{"comment":"The out-of-domain evaluation may be contaminated by training-data overlap. The Lizard dataset (Graham et al., 2021) is one of the six training datasets, and Lizard is known to be constructed by merging and re-annotating images from existing colorectal datasets, including CoNSeP (Graham et al., 2019), which is later listed among the 'remaining 8 datasets' used for out-of-domain evaluation. The manuscript does not report an image-level or slide-level de-duplication check between the Lizard training split and the CoNSeP test set. If any CoNSeP images or tiles appear in the Lizard training data, then CoNSeP is not 'not directly represented in the training set' as claimed, and the generalization result on CoNSeP is not valid evidence for the 'unseen dataset' claim. The authors must provide a de-duplication analysis (e.g., image hashing or patch matching) and re-run the out-of-domain evaluation excluding any overlapping images, or explicitly state which out-of-domain datasets remain genuinely unseen.","section":"Section 2.5 and Table 1"},{"comment":"All quantitative results are presented as figures with mean scores only; no error bars, standard deviations, or significance tests are reported. The claim that PathoSAM is 'the new state-of-the-art' for automatic nucleus instance segmentation rests on average scores over test sets, and for some comparisons the differences between variants (e.g., ViT-B vs. ViT-L vs. ViT-H) are described as small. Without uncertainty estimates or paired statistical tests on the same test images, the reader cannot assess whether the observed improvements over baselines are meaningful. Please provide numeric tables with per-dataset means and standard deviations (at least across test images), and perform significance tests for the headline comparisons that support the state-of-the-art claim.","section":"Section 3.1, Fig. 2a, Fig. 8, App. D.1"},{"comment":"The watershed seed thresholds used in AIS appear to be selected by a grid search on 15 validation images of LyNSeC, an out-of-domain dataset. The manuscript should clarify whether these thresholds were fixed before evaluation on all datasets and whether the LyNSeC validation images are disjoint from the LyNSeC test images. If the thresholds were selected on this out-of-domain validation split and then applied to the same dataset, the reported LyNSeC results are optimistically biased; if the thresholds were applied to other out-of-domain datasets, a sensitivity analysis should be reported to show that the main conclusions do not depend on the exact threshold choice.","section":"Fig. 10 and App. C"},{"comment":"The interactive segmentation comparison is limited to the original SAM and microSAM. The claim in Section 1 that 'PathoSAM outperforms other SAM variants for interactive segmentation' is broader than the evidence. Please either compare with additional SAM-based methods for histopathology (for example SAM-Path or MedicoSAM) or restrict the claim to the baselines actually evaluated, so that the stated claim matches the experimental scope.","section":"Section 3.2"}],"minor_comments":[{"comment":"The training iteration counts state '100,00 iterations' for the PanNuke and specialist models; this should read '100,000 iterations'.","section":"Appendix C"},{"comment":"The sentence 'We evaluate interactive segmentation (for SAM or variants) with AMG and, if available, AIS' is confusing because AMG and AIS are automatic segmentation methods, not interactive ones; it should be rephrased to refer to automatic segmentation.","section":"Section 2.2"},{"comment":"The notation I_P and I_B is used in the caption without a definition; it should be defined in the caption or explained at first use in the main text.","section":"Fig. 2 caption"},{"comment":"The phrase 'ca. 5 GB of VRAM per tile' should be clarified to mean peak VRAM usage during inference on a tile, not persistent per-tile allocation.","section":"Section 3.4"},{"comment":"The text refers to 'D.1 displays results' without identifying whether D.1 is a table or figure; the cross-reference should be made explicit.","section":"Appendix D.1"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the potential overlap between the Lizard training set and the CoNSeP out-of-domain test set; I would ask the editor to require an explicit de-duplication analysis before acceptance. In addition, the absence of numeric tables and uncertainty estimates makes the state-of-the-art claim difficult to verify. If the overlap is confirmed, the authors should re-run the out-of-domain evaluation without the contaminated dataset and revise the generalization claims accordingly. The paper is otherwise within scope and the open-source release is a positive feature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: PathoSAM is a solid, well-engineered extension of the µSAM recipe to histopathology, with a broad evaluation across 12 datasets and open-source code. The headline claim—new SOTA for automatic and interactive nucleus segmentation—is plausible but would be stronger with error bars and numeric tables. One specific leak risk needs checking: Lizard, a training dataset, is built in part from CoNSeP images, and CoNSeP is treated as out-of-domain. The paper never reports an image-level de-duplication check.\n\nThe genuinely new thing is joint training of interactive and automatic segmentation on a six-dataset H&E collection, evaluated on more datasets than prior SAM-for-histopathology work. The authors also test generalization to lymphocytes, glands, semantic segmentation on PanNuke, and a WSI pipeline. They are honest about where it fails: CryoNuSeg, lymphocyte-specific automatic segmentation, and semantic segmentation where CellViT still wins. The baselines are treated fairly—best available version per dataset—and the model is released.\n\nThe soft spots are real but not fatal. First, almost all quantitative results are in figures, not tables; the appendix figures are heatmaps and charts, so readers cannot extract exact numbers or confidence intervals. No error bars or significance tests anywhere. Second, the AIS watershed thresholds were chosen by grid search on 15 validation images of a single OOD dataset (LyNSeC); that is a tuned parameter on OOD data, so some OOD numbers are mildly optimistic. Third, the Lizard/CoNSeP overlap. Lizard is a merged dataset that includes CoNSeP images; CoNSeP is listed as OOD. If any CoNSeP images appear in the Lizard training split, the 'unseen dataset' claim for CoNSeP is contaminated. This is not a smoking gun—the other OOD datasets (NuInsSeg, CryoNuSeg, GlaS, etc.) are not affected—but the authors should check and report this. It matters because the paper's generalization claim is one of its main selling points.\n\nOverall, the paper is a worthwhile contribution. The core model and code are useful, and the evaluation, while imperfect, is honest and broad. The authors should be asked to add numeric tables and error bars, clarify the watershed threshold tuning, and run an image-level de-dup check between training and OOD sets. I would accept it for review, with the expectation of a revision.","headline":"A useful, well-engineered SAM-for-histopathology model with a broad evaluation; the SOTA claim is plausible but lacks error bars, and the Lizard/CoNSeP overlap needs checking before the OOD generalization claim can be trusted.","tokens_in":16577,"tokens_out":2615,"would_cite":true,"duration_ms":24569,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PathoSAM, a vision foundation model built by finetuning the Segment Anything Model on six H&E-stained nucleus datasets, claims state-of-the-art automatic and interactive nucleus instance segmentation in histopathology.","keywords":["segment anything","histopathology","nucleus instance segmentation","interactive segmentation","automatic segmentation","semantic segmentation","vision foundation model","whole-slide imaging"],"falsifier":"Evaluate PathoSAM on a new H&E dataset whose ground truth was drawn with a deliberately different nucleus convention, such as labeling cytoplasmic extensions or excluding small or overlapping nuclei, and compare automatic mean segmentation accuracy against a specialist model retrained on that dataset. If the generalist does not match or beat the specialist, or if errors concentrate wherever the annotation convention differs, the central generalization claim is falsified. A complementary check is to re-annotate CryoNuSeg to match PathoSAM's nucleus-region convention and see whether the reported quality gap disappears.","tokens_in":15613,"feed_emoji":"🔬","tokens_out":10424,"duration_ms":92526,"temperature":0.7,"pith_summary":"PathoSAM is a vision foundation model for nucleus segmentation in histopathology, trained by finetuning the Segment Anything Model on six H&E-stained nucleus datasets. The paper's central claim is that a single generalist model outperforms specialist models for automatic and interactive nucleus instance segmentation across 12 datasets, including out-of-domain data, and can be finetuned for semantic nucleus segmentation with strong results, though not above the current state-of-the-art CellViT. The intended upshot is that this removes the need to retrain per dataset: one set of weights supports interactive annotation and automatic segmentation on whole slides. The paper's own limitations state that gland and lymphocyte segmentation still require task-specific finetuning, and that automatic quality drops on CryoNuSeg where annotation conventions conflict with the visible nucleus region. If the claims hold, PathoSAM gives pathologists a practical single-model alternative to data-specific retraining.","feed_headline":"PathoSAM leads nucleus segmentation across 12 histology datasets","feed_subtitle":"Built by finetuning SAM on six H&E datasets, it handles automatic, interactive, and whole-slide nucleus segmentation.","key_machinery":"The load-bearing mechanism is joint training of SAM for both interaction and automation. PathoSAM retains SAM's image encoder, prompt encoder, and mask decoder for point, box, and mask prompting, and adds a convolutional decoder that outputs foreground probability, distance to the nearest object center, and distance to the nearest object boundary; seeded watershed converts those maps into instance segmentations. Training alternates between sampling prompts from ground-truth masks and correcting the model's predictions over seven iterations, while the same backbone is simultaneously optimized to reproduce the automatic-segmentation decoder's targets. At inference, the same weights answer a single click or box, refine a mask through further prompts, and run unattended over whole-slide tiles. A separate decoder added on top of the generalist performs semantic nucleus classification, trained with cross-entropy over PanNuke's five classes.","core_discovery":"On the paper's own terms, PathoSAM is the new state-of-the-art for automatic and interactive nucleus instance segmentation in histopathology. The model keeps SAM's prompt-driven mask prediction and adds a second decoder that predicts foreground probability and distance maps, which a seeded watershed turns into instance masks; the whole network is trained jointly with an objective that simulates interactive corrections and automatic segmentation on a diverse set of H&E nucleus datasets. In evaluation across 12 datasets, the ViT-L generalist achieves the highest average mean segmentation accuracy among automatic methods, even though for some competing methods the best available model version per dataset was used. For interactive segmentation, PathoSAM outperforms other SAM variants on single point and box prompts and reaches near-perfect accuracy after seven iterative corrections. The paper also reports a semantic segmentation decoder trained on PanNuke that ranks second behind CellViT, with the main shortfall on a rare 'dead cells' class, and a finetuning recipe that yields improved specialist models for lymphocyte and gland segmentation.","pith_inferences":["If annotation protocols were harmonized across datasets, the out-of-domain failures such as CryoNuSeg could shrink; the paper's own analysis attributes that failure to annotation inconsistency rather than image content alone.","The same joint interactive-plus-automatic training recipe may extend to multi-scale segmentation of glands and tissue regions if the model is given scale-aware prompts or multi-scale training, a step the paper identifies as future work.","Because PathoSAM's interactive segmentation is near-perfect after a few corrections, it could serve as a pre-annotation engine where human annotators confirm and fix masks instead of drawing nuclei from scratch.","A direct test of the semantic-segmentation gap would be to train with minority-class oversampling and joint instance-semantic objectives, which the paper proposes; if the gap to CellViT disappears, the current ranking is a training-strategy artifact rather than an architectural limit."],"forward_implications":["A single set of PathoSAM weights can replace dataset-specific retraining for nucleus instance segmentation across most H&E histopathology datasets.","Interactive annotation in tools such as QuPath and napari requires fewer prompts per nucleus, because point and box prompts already capture nuclei accurately.","Whole-slide images can be segmented automatically with tile-and-stitch scripts in under an hour on a GPU, enabling large-scale nucleus analysis.","The same finetuning recipe transfers to non-nucleus structures such as lymphocytes and glands when trained on task-specific data.","Semantic nucleus segmentation is possible from the same generalist backbone, though on PanNuke it trails CellViT mainly on a minority 'dead cells' class."],"supporting_citations":[{"why":"Supplies the Segment Anything Model architecture and prompt-driven interactive segmentation objective that PathoSAM finetunes.","marker":"Kirillov et al., 2023"},{"why":"Provides the joint training objective, the automatic instance segmentation decoder design, and the interactive evaluation protocol reused by PathoSAM.","marker":"Archit et al., 2025a"},{"why":"Defines the HoVerNet baseline and the CoNSeP dataset used in out-of-domain evaluation.","marker":"Graham et al., 2019"},{"why":"Contributes PanNuke, used for training a generalist, for the semantic segmentation benchmark, and for part of the evaluation.","marker":"Gamper et al., 2019"},{"why":"Supplies CellViT, the state-of-the-art semantic nucleus segmentation baseline that PathoSAM does not surpass.","marker":"Hörst et al., 2024"},{"why":"Defines the mean segmentation accuracy metric used to score all instance segmentation results.","marker":"Caicedo et al., 2019"},{"why":"Provides QuPath, one of the annotation tools in which PathoSAM was tested for interactive use.","marker":"Bankhead et al., 2017"}],"fun_headline_variants":["PathoSAM sets new bar for nucleus segmentation in pathology","Segment anything for histopathology? PathoSAM delivers","From SAM to PathoSAM: nucleus segmentation SOTA","PathoSAM: SOTA nucleus segmentation for histopathology"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the six H&E training datasets share consistent enough nucleus annotation conventions and enough imaging diversity for one model to generalize; the paper itself shows that when annotations disagree with the visible nucleus region, as in CryoNuSeg, automatic segmentation quality collapses.","fun_headline_variants_meta":{"raw":{"variants":["PathoSAM sets new bar for nucleus segmentation in pathology","Segment anything for histopathology? PathoSAM delivers","From SAM to PathoSAM: nucleus segmentation SOTA","PathoSAM: SOTA nucleus segmentation for histopathology"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000492,"raw_usage":{"total_tokens":2433,"prompt_tokens":974,"completion_tokens":1459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":1393}},"tokens_in":590,"tokens_out":1459,"duration_ms":10739,"temperature":1.0,"reasoning_tokens":1393,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T19:07:29.271569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate PathoSAM on a new H&E dataset whose ground truth was drawn with a deliberately different nucleus convention, such as labeling cytoplasmic extensions or excluding small or overlapping nuclei, and compare automatic mean segmentation accuracy against a specialist model retrained on that dataset. If the generalist does not match or beat the specialist, or if errors concentrate wherever the annotation convention differs, the central generalization claim is falsified. A complementary check is to re-annotate CryoNuSeg to match PathoSAM's nucleus-region convention and see whether the reported quality gap disappears.","supporting_citations":[{"cited_title":"Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick","cited_arxiv_id":null,"evidence_quote":"Supplies the Segment Anything Model architecture and prompt-driven interactive segmentation objective that PathoSAM finetunes."}],"review_version":1}