{"id":"d7216116-cc1d-46ca-9100-88fcc4688057","arxiv_id":"2411.18795","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new QuPath plugin uses five CircleNet models and weighted fusion to detect kidney glomeruli in whole slide images with a single click.","lead":"The paper presents GloFinder, a single-click QuPath plugin that detects kidney glomeruli across whole slide images using an ensemble of CircleNet models. It also reports faster manual correction times when clinicians edit the plugin's automatically generated annotations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported WCF gain may reflect threshold tuning to the 15 test WSIs: Section 3.2 describes T count=2 and T score=0.9 as 'carefully chosen' with no held-out validation, so the 0.829 mAP may be optimistically biased.","rationale":"The reader's weakest assumption correctly identifies the absence of a held-out validation set for the WCF thresholds as the key weakness. This is indeed the most load-bearing concern because the paper's primary quantitative claim is that WCF outperforms individual models and other ensemble methods. The thresholds T count=2 and T score=0.9 are the only hyperparameters of the fusion rule, and Section 3.2 describes them only as 'carefully chosen' without any description of a validation procedure. Given that all evaluation numbers in Table 1 come from the same 15 WSIs, the most parsimonious explanation for the 0.04 mAP gain could be selection on those 15 cases. The lack of error bars compounds this: with n=15 WSIs, the difference between 0.829 and 0.789 may well be within sampling noise. I do not see an internal inconsistency in the architecture or the experimental description; the concern is about the statistical validity of the reported gain. Because this issue requires the authors to provide additional validation evidence but does not by itself invalidate the tool, the conditional verdict remains appropriate. I would phrase the required condition as: report thresholds chosen on a validation split, or show through cross-validation that the reported advantage persists for fixed default thresholds, and provide per-WSI variability estimates.","tokens_in":5638,"tokens_out":2782,"duration_ms":26200,"concrete_test":"Hold out a random subset of, say, 5 of the 15 test WSIs, and optimize T count and T score on the remaining 10 WSIs (or on a separate validation set drawn from the training distribution). Freeze the chosen thresholds and evaluate GloFinder on the held-out 5 WSIs. Repeat over several random splits. If the held-out mAP(0.5:0.95) is no longer consistently above the best individual model (0.789), or if the difference falls within a bootstrap 95% confidence interval computed across the 15 WSIs, then the claimed superiority of WCF is not supported. Reporting per-split and bootstrap intervals would settle whether the 0.829 result is robust or an artifact of test-set threshold selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 states that WCF used 'carefully chosen thresholds': T count is set to 2 and T score to 0.9. No validation split is described anywhere in the manuscript; Section 3.1 says the 15 PAS-stained WSIs were used 'for testing and evaluation.' If these two fusion thresholds were selected by observing performance on those same 15 WSIs—the natural reading of 'carefully chosen' in a paper with no validation set—then the headline improvement of WCF (mAP 0.829) over the best single CircleNet model (mAP 0.789) is not an unbiased estimate of generalizable performance. The effect size is modest (0.04 mAP), the test set is small (15 WSIs, 2,051 glomeruli), and no error bars or per-WSI variance are reported, so even a mild amount of selection on the test set could explain the gap. The central claim that WCF achieves superior detection performance therefore rests on the assumption that the fusion hyperparameters are fixed a priori or tuned on independent data. That assumption is not supported by the manuscript as written. A secondary issue is that the code link in the abstract points to hrlblab/GloFinder while the 'Code and Demo Video' section points to hrlblab/PathVisual; this inconsistency prevents independent verification, but the threshold-selection concern is the more load-bearing problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GloFinder, a QuPath plugin for whole-slide-image (WSI) glomerular detection that runs five CircleNet models in the background and combines their outputs using a proposed Weighted Circle Fusion (WCF) ensemble. The plugin is designed for single-click operation in QuPath, with graphical visualization and editing of detected circles. The authors report that WCF achieves mAP(0.5:0.95)=0.829 on 15 PAS-stained WSIs (2,051 glomeruli), outperforming the best individual CircleNet model (0.789) and standard NMS/Soft-NMS baselines. They further report a human-in-the-loop annotation study in which annotation time drops from 9.23 to 2.9 minutes per WSI, a 68.59% reduction. The paper's contributions are framed as (1) an accessible QuPath plugin, (2) improved detection via WCF, and (3) an efficient annotation workflow.","tokens_in":6086,"tokens_out":3614,"duration_ms":32553,"significance":"If the reported gains are robust, GloFinder addresses a real usability gap in computational pathology: it turns a research detection model into a clinician-accessible GUI tool with WSI-level processing and online editing, trained on a substantial corpus of ~160,000 manually annotated glomeruli. The open-source plugin, the single-click workflow, and the human-in-the-loop curation concept are all practical contributions that could benefit nephropathology practice. However, the current evidence base is thin: the central detection claim rests on a single table with point estimates from 15 WSIs, no error bars or significance tests, and WCF hyperparameters that appear to be selected without a held-out validation set. The annotation-time reduction also lacks protocol detail. These issues do not invalidate the engineering contribution, but they do mean the quantitative superiority claims are not yet established to the standard expected for a peer-reviewed venue.","major_comments":[{"comment":"The WCF thresholds T_count=2 and T_score=0.9 are described as 'carefully chosen,' but the manuscript describes no validation split: Section 3.1 states that the 15 PAS-stained WSIs were used 'for testing and evaluation.' If these thresholds were tuned on the same 15 WSIs, the reported mAP of 0.829 is optimistically biased, and the 0.040 mAP improvement over the best single model (0.789) may partly reflect test-set selection rather than a generalizable ensemble gain. Please specify how the thresholds were selected, report the validation data and results if they exist, or provide a sensitivity analysis over the threshold grid showing that the reported operating point is not a cherry-picked optimum.","section":"Section 3.2 and Table 1"},{"comment":"All detection comparisons are reported as single point estimates without error bars, per-WSI variance, or significance tests. With only 15 test WSIs, the difference between WCF and the best single CircleNet (0.040 mAP) could easily fall within noise. The claim that WCF achieves 'significantly higher' mAP values requires per-WSI mAP distributions, bootstrap confidence intervals, and a paired significance test (e.g., Wilcoxon signed-rank) comparing WCF with each individual model and with NMS/Soft-NMS. Without this, the central detection improvement is not statistically substantiated.","section":"Table 1 and Section 4.1"},{"comment":"The human-in-the-loop time study lacks the protocol detail needed to support the 68.59% reduction claim. The manuscript does not state how many annotators participated, their experience level, the instructions given, how manual annotation time was measured, what correction operations were counted, or how the 10 WSIs were selected. It also reports only mean times without per-WSI paired comparisons or confidence intervals. Please provide a complete protocol and report the paired per-WSI times with appropriate statistics; otherwise the annotation-efficiency contribution is unverifiable.","section":"Section 4.2"},{"comment":"The abstract states that code and the QuPath plugin are available at https://github.com/hrlblab/GloFinder, but the 'Code and Demo Video' section gives https://github.com/hrlblab/PathVisual. This inconsistency prevents independent verification of the implementation and should be resolved: the manuscript must point to a single, accessible repository containing the plugin, the models, and the evaluation code, or explain the relationship between the two repositories.","section":"Code and Demo Video section; Abstract"}],"minor_comments":[{"comment":"The phrase 'a AI-empowered QuPath plugin' should be 'an AI-empowered QuPath plugin.'","section":"Section 1"},{"comment":"The caption refers to red and blue highlighting for the highest and second-highest values, but the table as rendered in the text may not show color; please add symbolic markers (e.g., superscripts or bold/underline) so the distinctions are visible in grayscale print.","section":"Table 1 caption"},{"comment":"The role of NMS is ambiguous: the text says NMS filters redundant detections before fusion, but it also appears in Table 1 as a standalone ensemble method. Please clarify whether NMS is applied per-model before WCF, after WCF, or both, and define the exact NMS variant and IoU threshold used.","section":"Section 2.2"},{"comment":"The statement that 'approximately 30,000 glomeruli varied between the training datasets of different models' is vague; please specify the total number of unique glomeruli, the size of the intersection/union across the five training sets, and how the varied subsets were constructed.","section":"Section 3.1"},{"comment":"The Discussion lists 'detection time required to process full WSIs' as a limitation, but Section 3.4 reports an average of 21 seconds per WSI; please contextualize whether this is the full WSI processing time or a per-patch time, and clarify what hardware assumption applies.","section":"Section 5"},{"comment":"The term 'state-of-the-art (SOTA) CircleNet method' is used without comparison to other recent glomerular detection or object detection approaches; please soften the claim or add a brief comparison to justify the SOTA designation.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The engineering contribution is real and potentially valuable, but the quantitative claims need substantial additional evidence. The most important risk is the WCF threshold-selection issue, because it directly affects the headline detection improvement; a clean validation procedure or sensitivity analysis would materially change my assessment. The code-link inconsistency should also be fixed before publication, as it currently blocks reproducibility. I recommend major revision rather than rejection because the central ideas are sound and the issues are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GloFinder is a useful, well-motivated software paper that is best read as a tool contribution, not a new detection algorithm. The plugin itself—single-click CircleNet inference across whole slides inside QuPath, with color-coded output and online editing—is a real answer to an actual accessibility problem, and the HITL time saving (9.23 to 2.9 min per image) is plausible and valuable. The paper deserves a serious referee.\n\nWhat is genuinely new: the QuPath integration, the GUI workflow, and the human-in-the-loop evaluation. The underlying CircleNet and WCF are prior work, but repackaging them into a clinician-usable plugin is a legitimate engineering contribution. The training scale (160k glomeruli) is decent, and the test set of 15 WSIs is independent, though small.\n\nThe soft spots are the ones you flagged. Section 3.2 describes the WCF thresholds as \"carefully chosen\" with no held-out validation split, and the 15 WSIs are the only test data described. So the 0.829 vs 0.789 mAP gap may partly reflect tuning to the test set. The lack of error bars or per-WSI variance makes the 0.04 mAP difference impossible to assess. This is a genuine weakness, but it doesn't sink the paper's main purpose: the plugin would still help even if WCF's advantage over a single model is smaller than claimed. The HITL study is also under-specified—no number of annotators, no inter-operator variance, no detail on what \"editing\" involved. And the code link mismatch (abstract points to GloFinder repo, Code section points to PathVisual) is sloppy and should be fixed before publication.\n\nI don't think this is overclaiming in bad faith; it reads like a tools paper that didn't take the extra step of a proper validation split. The authors should either add validation-based threshold selection, freeze thresholds a priori, or openly state that the thresholds are tuned and the reported gain is optimistic. The citation to their own WCF preprint is appropriate since they are using that method.\n\nWho benefits: nephropathology researchers, clinicians wanting off-the-shelf detection, and anyone building QuPath plugins. It is worth engaging with. Send it to peer review with the expectation of a revision addressing the threshold selection, error bars, and code link. A desk rejection would be a miss.","headline":"A genuinely useful QuPath plugin for glomerular detection, but the validation section overreaches: the WCF advantage may be threshold tuning on the test set, and the paper needs a proper validation split before the performance claims are publishable.","tokens_in":6484,"tokens_out":2959,"would_cite":false,"duration_ms":27322,"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":"GloFinder packages automated glomerular detection as a single-click QuPath plugin and claims that ensembling five CircleNet models with Weighted Circle Fusion improves whole-slide detection accuracy over any single model.","keywords":["Automated Glomeruli Detection","Whole Slide Images","CircleNet","Weighted Circle Fusion","QuPath Plugin","Renal Pathology","Medical Image Analysis","Human-in-the-loop Annotation"],"falsifier":"Select fusion thresholds on a distinct validation set, then measure $mAP(0.5{:}0.95)$ on an untouched test set; if the ensembled detector no longer beats the best single CircleNet model, the reported advantage is an artifact of threshold selection.","tokens_in":5472,"feed_emoji":"🔬","tokens_out":5061,"duration_ms":35726,"temperature":0.7,"pith_summary":"GloFinder packages automated glomerulus detection as a QuPath plugin so that pathologists and clinicians can run whole-slide analysis with a single click and without writing code. The paper's central claim is that ensembling five CircleNet models with Weighted Circle Fusion (WCF) improves detection accuracy over any single model, reaching $mAP(0.5{:}0.95)=0.829$ on 15 PAS-stained mouse kidney WSIs compared with the best single model's 0.789. It further claims the plugin's human-in-the-loop workflow cuts annotation time from 9.23 to 2.9 minutes per image, a 68.59% reduction. If these claims hold, AI-assisted kidney pathology becomes practical for non-programmers and large-scale curation of glomerular annotations becomes faster.","feed_headline":"One-click plugin detects glomeruli, cuts annotation time 69%","feed_subtitle":"GloFinder fuses five CircleNet models in QuPath, reaching mAP 0.829 and enabling pathologists to edit results directly.","key_machinery":"The central object is the circle representation: CircleNet predicts each glomerulus as a circle (center coordinates plus radius), a rotation-consistent representation suited to circular objects. Weighted Circle Fusion (WCF) is the ensemble mechanism that aggregates overlapping circles from multiple CircleNet models by weighting and merging based on confidence scores, using count and score thresholds to suppress low-consensus detections. The plugin's pipeline tiles the WSI into half-overlapping patches, runs five models, transforms coordinates back to whole-slide space, applies NMS, then WCF, and displays fusion results with colors indicating how many models agreed.","core_discovery":"The paper discovers that fusing the circle predictions of five independently trained CircleNet models, each trained on more than 160,000 manually annotated mouse glomeruli, yields a detector that outperforms every individual model and other fusion baselines such as NMS and Soft-NMS on an independent test set of 15 PAS-stained WSIs. The WCF ensemble weights overlapping detected circles by confidence and requires consensus from at least two models (T count = 2) with a high score threshold (T score = 0.9). When deployed inside the QuPath interface with color-coded visualization of consensus, GloFinder achieves the reported accuracy while cutting annotation time by roughly two-thirds in a human-in-the-loop setup.","pith_inferences":["The reported 0.829 mAP likely depends on the fusion thresholds (T count = 2, T score = 0.9) being tuned on the same test set; a held-out validation split would reveal whether the ensemble advantage persists.","A testable extension is to evaluate WCF on other organ systems with circular objects and measure whether the consensus rule reduces annotation effort proportionally.","Cloud-based execution could remove the local GPU requirement that currently limits accessibility for under-resourced labs.","The 21-second per-slide runtime suggests the bottleneck for clinical adoption is not detection speed but integration with existing pathology information systems."],"forward_implications":["If GloFinder's claims are correct, non-programmer pathologists can perform WSI-level glomerular detection in QuPath at 21 seconds per slide.","Color-coded consensus visualization lets clinicians prioritize review of low-consensus circles, which are more likely to be errors.","The same plugin architecture can be retargeted to other circular biomedical objects, such as cell nuclei, by swapping the trained model.","Human-in-the-loop annotation becomes fast enough to support large-scale curation of training data for future models.","Ensembling models trained on diverse datasets through WCF may generalize better to unseen whole-slide images than any single model."],"supporting_citations":[{"why":"Supplies the CircleNet detector architecture and the circle IoU metric used for evaluation.","marker":"[1]"},{"why":"Provides the QuPath open-source platform that the plugin integrates with.","marker":"[2]"},{"why":"Establishes the Weighted Circle Fusion ensemble method that combines the five models' outputs.","marker":"[3]"},{"why":"Supplies the NMS algorithm used for redundancy removal and as a baseline fusion comparison.","marker":"[5]"},{"why":"Provides the Soft-NMS baseline that WCF must outperform.","marker":"[6]"}],"fun_headline_variants":["Five-model AI fusion spots kidney glomeruli in one click","GloFinder: one-click glomeruli detection, 69% faster annotation","AI plugin fuses models to find glomeruli at mAP 0.829","Open-source QuPath plugin detects glomeruli, cuts annotation time 69%","Single-click glomeruli detection with AI ensemble in QuPath"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fusion thresholds were not validated on a separate set; they were 'carefully chosen' and applied to the same 15 WSIs used for evaluation, so the measured gain over single models might partly reflect threshold tuning rather than a general property of the fusion method.","fun_headline_variants_meta":{"raw":{"variants":["Five-model AI fusion spots kidney glomeruli in one click","GloFinder: one-click glomeruli detection, 69% faster annotation","AI plugin fuses models to find glomeruli at mAP 0.829","Open-source QuPath plugin detects glomeruli, cuts annotation time 69%","Single-click glomeruli detection with AI ensemble in QuPath"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000446,"raw_usage":{"total_tokens":2247,"prompt_tokens":933,"completion_tokens":1314,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1220}},"tokens_in":549,"tokens_out":1314,"duration_ms":9558,"temperature":1.0,"reasoning_tokens":1220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:51:51.222994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Select fusion thresholds on a distinct validation set, then measure $mAP(0.5{:}0.95)$ on an untouched test set; if the ensembled detector no longer beats the best single CircleNet model, the reported advantage is an artifact of threshold selection.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CircleNet detector architecture and the circle IoU metric used for evaluation."},{"cited_title":"Bankhead, M","cited_arxiv_id":null,"evidence_quote":"Provides the QuPath open-source platform that the plugin integrates with."},{"cited_title":"Neubeck and L","cited_arxiv_id":null,"evidence_quote":"Supplies the NMS algorithm used for redundancy removal and as a baseline fusion comparison."},{"cited_title":"Bodla, B","cited_arxiv_id":null,"evidence_quote":"Provides the Soft-NMS baseline that WCF must outperform."}],"review_version":1}