{"id":"09f8bad5-2e83-44ce-9ca6-ee20eca3f89d","arxiv_id":"2411.15514","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CellPilot, a SAM-based model fine-tuned with LoRA on nine histopathology datasets, achieves better interactive segmentation than SAM, MedSAM, and SimpleClick on most held-out cell and gland datasets.","lead":"CellPilot combines a fine-tuned version of the Segment Anything Model with CellViT to segment cells and glands in histopathology images, supporting both automatic and interactive refinement. The authors show it beats three existing interactive tools on three held-out datasets, with the code and a graphical interface released as open source.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 contradicts the abstract's claim of 'superior performance compared to other interactive tools on three held-out datasets': on CRAG with a single point prompt, SAM (0.35) exceeds CellPilot (0.32), making the central claim internally inconsistent.","rationale":"The reader correctly noted that the abstract overstates results on CRAG, but the reader's designated weakest assumption was the realism of the simulated interactive refinement protocol. The present stress test identifies a more direct and falsifiable flaw: the paper's own Table 2 shows SAM outperforming CellPilot on CRAG with a single point prompt (0.35 vs. 0.32), directly contradicting the abstract's claim of superior performance on three held-out datasets. This is not a matter of user-behavior simulation; it is an internal inconsistency in the reported numbers. The paper still presents useful contributions (large-scale training, open-source code, strong results on cell datasets and box prompts), so a conditional acceptance with mandatory revision is appropriate rather than rejection. The authors should correct the overbroad claim, report paired statistics or confidence intervals where the means are close, and, ideally, evaluate the CellViT-based automatic segmentation pathway that is described but never assessed. This agrees partially with the reader because the reader mentioned the abstract overstatement in the rationale but did not make it the central load-bearing concern.","tokens_in":5745,"tokens_out":3435,"duration_ms":30714,"concrete_test":"Reproduce the CRAG single-point-prompt row of Table 2 using the released code and the protocol in Section 2.3 (resize so the long side is 1024, pad to 1024×1024, select ten random masks per image, sample one random interior point per mask, use the same seeds). Compute mean IoU for CellPilot and SAM, and also compute a paired difference (CellPilot − SAM) with bootstrap confidence intervals. If CellPilot is not significantly better than SAM, the claim 'superior performance on three held-out datasets' must be revised to exclude single-point gland segmentation; if CellPilot is actually better, then Table 2 contains a numerical error that should be corrected.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim, stated in the abstract and reiterated in the conclusion, is that CellPilot surpasses SAM, MedSAM, and SimpleClick across the three held-out datasets. This claim is contradicted by the paper's own Table 2 (Section 2.4): on CRAG with a single point prompt, SAM achieves a mean IoU of 0.35±0.33 while CellPilot achieves 0.32±0.33. The text even acknowledges that 'using a point prompt, all models perform weak, with SAM having a slight advantage over CellPilot,' but this qualification is absent from the abstract and ignored in the conclusion's statement that CellPilot 'surpasses SAM, MedSAM, and SimpleClick across three different held-out datasets.' Because the reported numbers themselves falsify the headline comparison for one dataset and one prompt modality, the main empirical claim is overstated as written. This is not a matter of external consensus or statistical subtlety; it is an internal inconsistency. A secondary but related issue is that the automatic segmentation component via CellViT boxes is never evaluated, so the claim of 'enabling automatic segmentation' remains unverified; however, the CRAG single-point result already requires a correction to the paper's primary claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CellPilot, a unified framework for automatic and interactive segmentation of cells and glands in histopathology. CellPilot builds on SAM by fine-tuning the prompt encoder and mask decoder and applying LoRA to the image encoder, and it uses CellViT to propose bounding boxes for automatic initialization. The model is trained on more than 675,000 masks from nine datasets spanning 16 organs and is evaluated against SAM, MedSAM, and SimpleClick on three held-out datasets (CellSeg, MoNuSAC, CRAG) under a simulated interactive refinement protocol with point and box prompts. The authors report higher mean IoU for CellPilot in most settings, plus a graphical user interface for dataset creation, and they release the model open-source.","tokens_in":6007,"tokens_out":3853,"duration_ms":34796,"significance":"If the performance claims hold, CellPilot would be a useful contribution to histopathology segmentation because it combines automatic initialization with interactive refinement in a single framework, is trained on an unusually large and diverse set of histopathology masks, and ships as an open-source model with a GUI. The use of the established SimpleClick simulation protocol and the inclusion of SAM, MedSAM, and SimpleClick as baselines are strengths. However, the central claim of superior performance is overstated relative to the paper's own Table 2, and the automatic segmentation component is not evaluated at all. Because these issues affect the headline claims rather than just presentation, the manuscript needs substantive revision before it can be accepted.","major_comments":[{"comment":"The abstract states that CellPilot \"demonstrates superior performance compared to other interactive tools on three held-out histopathological datasets,\" and the conclusion repeats that the model \"surpasses SAM, MedSAM, and SimpleClick across three different held-out datasets.\" This is internally inconsistent with Table 2, where on CRAG with a single point prompt SAM achieves mean IoU 0.35±0.33 while CellPilot achieves 0.32±0.33. Section 2.4 itself qualifies the result by saying that \"using a point prompt, all models perform weak, with SAM having a slight advantage over CellPilot.\" The abstract and conclusion should be reworded to specify the settings in which CellPilot is superior, e.g., box prompts and multi-point refinement, or \"all settings except single-point gland prompts.\"","section":"Abstract and Conclusion vs. Section 2.4, Table 2"},{"comment":"The automatic segmentation stage is a headline feature of the framework but is never evaluated. Section 2.1 describes CellViT generating bounding boxes that serve as initial prompts for cells, yet no experiment reports the quality of these automatic predictions, such as the IoU of CellViT-initialized masks before interactive refinement or a comparison against running CellViT alone. Consequently, the claim of \"enabling automatic segmentation\" is unverified. The paper should either include an evaluation of the automatic stage or explicitly scope the contributions to interactive refinement only.","section":"Section 2.1 and Figure 1"},{"comment":"No statistical significance testing or confidence intervals are reported for the comparative results. The standard deviations in Table 2 are large, e.g., ±0.33 for both SAM and CellPilot on CRAG with a point prompt, and several reported differences, such as CellPilot 0.63±0.27 versus SAM 0.60±0.31 on CellSeg with a point prompt, are within one standard deviation. Figure 2 shows mean curves without error bars. Paired significance tests or confidence intervals across the same masks/images are needed to support the repeated claims that CellPilot \"consistently improves\" and outperforms the baselines.","section":"Section 2.3 and Section 2.4"},{"comment":"The interactive refinement protocol used for evaluation is the same simulated click protocol used for training: random initial points or boxes, followed by iterative sampling from the largest error region. While this protocol is standard and is applied identically to all baselines, the paper does not acknowledge that real-user click behavior may differ from this simulation, which could affect the transferability of the interactive comparisons. An explicit limitation statement, or a sensitivity experiment with an alternative prompt-sampling strategy, would strengthen the evaluation.","section":"Section 2.1 versus Section 2.3"}],"minor_comments":[{"comment":"The caption says \"totaling over 9,900 images,\" but the listed numbers sum to 7,913 training images and 2,074 test images; please state the train/test split explicitly to avoid ambiguity.","section":"Table 1 caption"},{"comment":"The name \"MedSAM\" is spelled inconsistently as \"MedSam\" in Figure 2 and its caption; please use one spelling throughout.","section":"Table 2 and figure texts"},{"comment":"The individual panels are dense and the curves are hard to distinguish when printed in grayscale; adding per-panel line labels or distinct marker styles would improve readability.","section":"Figure 2"},{"comment":"The sentence \"CellPilot achieves high mean IoU scores using a single point or box prompt in cell segmentation tasks\" is correct, but the following sentence about gland segmentation should appear immediately after the CRAG numbers so the reader is not misled by the earlier results.","section":"Section 2.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is promising and the resource release is valuable, but the mismatch between the abstract/conclusion and the CRAG single-point result in Table 2 is a clear internal inconsistency that must be fixed. The missing evaluation of the automatic CellViT-initialization stage is also a substantive gap relative to the paper's stated scope. With those corrected, the manuscript could be suitable; I would not reject it outright, but the authors need to revise the claims and add the missing experiments or scoping statements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CellPilot is worth a look if you care about interactive segmentation for pathology, but the paper's headline claim is internally inconsistent. On CRAG with a single point prompt, Table 2 has SAM at 0.35 mean IoU and CellPilot at 0.32. That directly contradicts the abstract's 'superior performance compared to other interactive tools on three held-out datasets.' The text itself admits 'SAM having a slight advantage,' yet the abstract and conclusion don't.\n\nWhat is genuinely new: the combination of CellViT boxes as automatic proposals and a LoRA-fine-tuned SAM for interactive point/box refinement. The scale is real—9 training datasets, 675k masks, 16 organs—and the code and GUI are open-source. That is a practical contribution for building annotated histopathology datasets. Training follows SimpleClick's simulated correction protocol, which is sensible and standard.\n\nThe soft spots are in the evaluation and the claims. First, the CRAG point-prompt result is not a subtle statistical quibble; the numbers in the paper falsify the blanket claim. Second, the 'automatic segmentation' half of the framework is never evaluated. CellViT boxes are only described as a way to start the interactive loop; no quantitative result shows the automatic stage working. So 'enabling automatic segmentation' is unverified. Third, there are no significance tests, and the standard deviations are large (e.g., 0.32±0.33). The qualitative ranking may well be right, but the paper doesn't show it. Fourth, the simulation trains and evaluates on the same click protocol. That is fair to baselines because they all run under the same simulator, but it does not tell you how any of these models behave with real users. I'd call that a minor caveat, not a fatal flaw.\n\nWho is this for? Researchers building interactive segmentation tools for pathology and practitioners who want to annotate large datasets. It deserves a serious referee. The engineering is solid and the model is useful, but a revision needs to correct the abstract and conclusion, report stats or at least acknowledge the overlap, and ideally evaluate the automatic stage. I would not reject it; I would send it back for major revision.\n\nSo my recommendation: engage with it, but hold the authors to their own table.","headline":"Useful interactive segmentation tool for pathology, but the abstract overclaims: Table 2 itself shows SAM beating CellPilot on CRAG with a single point prompt.","tokens_in":6486,"tokens_out":2221,"would_cite":true,"duration_ms":19254,"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":"The paper claims that fine-tuning SAM on 675,000 histopathology masks yields a single model that beats SAM, MedSAM, and SimpleClick under simulated interactive refinement on three held-out datasets, while also enabling automatic…","keywords":["cell segmentation","gland segmentation","interactive segmentation","histopathology","Segment Anything Model","LoRA fine-tuning","automatic segmentation"],"falsifier":"Ask pathologists to refine CellPilot and SAM outputs with their own free-form clicks and compare mean IoU gained per click; if CellPilot is not ahead, the simulation-based claim is falsified. Separately, feed the automatic CellViT boxes through CellPilot and compare with CellViT alone to test the unverified automatic stage.","tokens_in":5580,"feed_emoji":"🔬","tokens_out":8978,"duration_ms":70700,"temperature":0.7,"pith_summary":"The paper tries to establish that one model can handle both automatic and interactive cell and gland segmentation in histopathology, so a practitioner can start with an automatic mask and fix it with a few clicks instead of switching tools. The authors fine-tune SAM on nine public datasets covering 16 organs and more than 675,000 masks, using a simulated click protocol that mimics corrections. On three held-out datasets they report higher mean IoU than SAM, MedSAM, and SimpleClick for cell segmentation with point or box prompts and for gland segmentation with box prompts; the one exception is gland segmentation from a single point, where SAM keeps a small lead. If these results hold, the practical payoff is a faster path to large annotated histopathology datasets and a single tool for automatic and interactive analysis.","feed_headline":"CellPilot tops SAM, MedSAM, SimpleClick in histopathology segmentation","feed_subtitle":"One point or box prompt refines cells and glands across three held-out datasets; open-source GUI included.","key_machinery":"The central machinery is a fine-tuned Segment Anything Model (SAM): its image encoder is adapted to histopathology with low-rank adaptation (LoRA), while its prompt encoder and mask decoder are retrained on the nine-dataset cell and gland collection. The prompt simulation of SimpleClick, which starts from a point or box and repeatedly adds positive or negative points in the largest error region of the current prediction, teaches the model to correct itself during training. CellViT supplies the initial bounding boxes that turn the same interactive model into an automatic segmenter for cells.","core_discovery":"CellPilot is a two-stage framework: CellViT proposes bounding boxes for cells, and a fine-tuned SAM decodes them into masks; a user then adds positive or negative point or box prompts to correct errors. The authors' central discovery is that this combined model, trained on nine cell and gland datasets with more than 675,000 masks across 16 organs, generalizes to three held-out datasets and reports higher mean IoU than SAM, MedSAM, and SimpleClick for interactive cell segmentation with both point and box prompts, and for interactive gland segmentation with box prompts. The exception is single-point gland segmentation on CRAG, where SAM retains a small lead; MedSAM fails badly with point prompts. The same network, prompted by CellViT boxes, is intended to provide automatic segmentation, though that automatic stage is not separately evaluated.","pith_inferences":["A real-user study is the natural next test: the simulation assumes users click in the largest error region, but pathologists may click on boundaries or small structures, and the reported advantage could shrink or grow depending on that behavior.","The automatic stage is never evaluated in the paper, so the claim of one framework for both automatic and interactive segmentation is only half-supported; comparing CellPilot's CellViT-box initializations against CellViT alone would settle it.","Because the model has no explicit cell-versus-gland class token, confusion between the two object types is plausible; adding a lightweight classifier, as the paper's conclusion suggests, could fix this and enable semantic cell typing.","A practical extension is to use the GUI to collect logged user interactions, which would let the community replace the simulated refinement protocol with a real click distribution and retrain on it."],"forward_implications":["A single model can replace separate automatic and interactive tools for cell and gland segmentation, so a correction made by hand also improves the same network that produced the initial mask.","The training protocol, which simulates corrective clicks from the largest error region, scales to large multi-dataset collections without manual annotation effort, which is why the model can be trained on 675,000 masks.","For cell segmentation, the reported mean IoU advantage holds for both single-point and single-box prompts on held-out datasets, and successive corrective points keep improving the masks.","For gland segmentation with a box prompt, the model reports the highest mean IoU among the compared tools, while single-point gland prompts remain a weak spot for every method.","The open-source GUI turns the model into an annotation accelerator, allowing practitioners to build large histopathology datasets with less manual tracing."],"supporting_citations":[{"why":"It supplies the CellViT model that generates initial bounding boxes for automatic cell segmentation.","marker":"[1]"},{"why":"SAM is the base model being fine-tuned and the strongest interactive baseline in the comparison.","marker":"[2]"},{"why":"MedSAM is the medical-imaging adaptation used as a baseline, and its box-only behavior explains its failure with point prompts.","marker":"[4]"},{"why":"SimpleClick supplies both the baseline and the simulated interactive refinement protocol, with prompts sampled from the largest error region, used for training and evaluation.","marker":"[8]"},{"why":"CellSeg is a held-out multimodal cell segmentation test set used in the comparison.","marker":"[21]"},{"why":"MoNuSAC is a held-out multi-organ nuclei segmentation test set used in the comparison.","marker":"[22]"},{"why":"CRAG is the held-out gland segmentation test set used in the comparison.","marker":"[23]"}],"fun_headline_variants":["CellPilot unifies auto and interactive histopathology segmentation","CellPilot beats SAM, MedSAM, SimpleClick in interactive histo","One point or box prompt: CellPilot refines cells and glands","CellPilot trained on 675k masks, 16 organs, open-source GUI","CellPilot: open-source GUI for interactive cell and gland segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that simulated corrective clicks, always placed in the largest wrong region, match how a real user would correct a segmentation, and that the automatic starting boxes actually work.","fun_headline_variants_meta":{"raw":{"variants":["CellPilot unifies auto and interactive histopathology segmentation","CellPilot beats SAM, MedSAM, SimpleClick in interactive histo","One point or box prompt: CellPilot refines cells and glands","CellPilot trained on 675k masks, 16 organs, open-source GUI","CellPilot: open-source GUI for interactive cell and gland segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001741,"raw_usage":{"total_tokens":6849,"prompt_tokens":883,"completion_tokens":5966,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":5870}},"tokens_in":499,"tokens_out":5966,"duration_ms":37566,"temperature":1.0,"reasoning_tokens":5870,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:11:38.142519+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ask pathologists to refine CellPilot and SAM outputs with their own free-form clicks and compare mean IoU gained per click; if CellPilot is not ahead, the simulation-based claim is falsified. Separately, feed the automatic CellViT boxes through CellPilot and compare with CellViT alone to test the unverified automatic stage.","supporting_citations":[{"cited_title":"The seg- mentation of cells and glands in tissue allows for the analy- sis of shape and frequency, which can be indicators of dis- ease","cited_arxiv_id":null,"evidence_quote":"It supplies the CellViT model that generates initial bounding boxes for automatic cell segmentation."},{"cited_title":"Method The CellPilot framework consists of a combination of a fine- tuned version of SAM and CellViT (figure 1a)","cited_arxiv_id":null,"evidence_quote":"SAM is the base model being fine-tuned and the strongest interactive baseline in the comparison."},{"cited_title":"Ethical ap- proval was not required as confirmed by the license attached with the open access data","cited_arxiv_id":null,"evidence_quote":"MedSAM is the medical-imaging adaptation used as a baseline, and its box-only behavior explains its failure with point prompts."},{"cited_title":"Cellvit: Vision transformers for precise cell segmentation and classification,","cited_arxiv_id":null,"evidence_quote":"SimpleClick supplies both the baseline and the simulated interactive refinement protocol, with prompts sampled from the largest error region, used for training and evaluation."},{"cited_title":"Segpc-2021: A challenge & dataset on segmentation of multiple myeloma plasma cells from microscopic images,","cited_arxiv_id":null,"evidence_quote":"CellSeg is a held-out multimodal cell segmentation test set used in the comparison."},{"cited_title":"Gland segmentation in colon histology images: The glas challenge contest,","cited_arxiv_id":null,"evidence_quote":"MoNuSAC is a held-out multi-organ nuclei segmentation test set used in the comparison."},{"cited_title":"Use Case 1: Nuclei Segmentation - Andrew Janowczyk,","cited_arxiv_id":null,"evidence_quote":"CRAG is the held-out gland segmentation test set used in the comparison."}],"review_version":1}