{"id":"1ea153e2-7e21-47ed-ad2a-c8ab33c50a39","arxiv_id":"2411.18893","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Convex hull post-processing raises kidney glomeruli segmentation Dice scores by 0.005 to 0.033 across four models and four data splits.","lead":"This paper proposes CovHuSeg, a post-processing step that replaces a deep learning segmentation mask with the convex hull of its contour, and tests it on kidney glomeruli segmentation. The authors report consistent Dice score improvements across four models, with larger relative gains on noisy and limited-data settings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run Dice gains of 0.005–0.033 lack statistical support; retraining with multiple seeds is needed to confirm the claimed universal improvement from CovHuSeg.","rationale":"I read the paper in good faith. The algorithm is simple and clearly described, and the reported direction of the effect is consistent with the convexity rationale: if ground-truth glomeruli are largely convex and model predictions have boundary concavities or holes, then convex-hull filling will improve Dice. The reader's weakest_assumption (convexity) is a plausible mechanistic premise. However, the central claim is an empirical universal improvement, and the evidence for it rests entirely on single training runs. The smallest reported gain (0.005 Dice) is far below what one would expect from seed variation alone, and the paper gives no distributional data. I therefore see the absence of statistical validation as the most load-bearing concern: if the gains vanish under reseeding, the central claim collapses, regardless of the convexity assumption. My proposed test directly targets this. The reader's rationale already lists missing error bars as a condition, so my concern partially overlaps; I differ in weighting it as the primary risk rather than convexity. Verdict remains CONDITIONAL because the method is plausible and the tables are suggestive, but the empirical claim is not yet robust.","tokens_in":8433,"tokens_out":5735,"duration_ms":57524,"concrete_test":"Retrain each of the four models on each of the four splits with at least five different random seeds (40 runs total), evaluate each trained model with and without CovHuSeg on the same test set, and report the mean and 95% confidence interval of the Dice difference per split and noise condition. Perform a paired Wilcoxon signed-rank test across the 2,305 test patches (or across the 12 WSIs) for each configuration. If the lower confidence bound of the Dice difference includes zero in more than a small minority of configurations, or if the paired test is not significant at p<0.05, then the universal-improvement claim is not established. As a supplementary check, also compute the Dice change per connected component to verify that CovHuSeg never merges distinct glomeruli.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim (Abstract; Section 4.3) is that CovHuSeg increases Dice for all four models across all four data splits on both normal and noisy test sets. However, Tables 1 and 2 report a single Dice value per configuration, with no error bars, no multiple seeds, and no paired statistical test. Deep-learning training is stochastic, and the reported absolute gains are as small as 0.005 Dice for UNet (e.g., Table 1, Split B: 0.858 to 0.863) and 0.008 (Table 2, Split B). These increments are within the typical run-to-run variation of segmentation models trained for 150 epochs, so the observed improvements could be seed luck rather than a real effect of the post-processing. The convexity rationale in Section 3.2 would predict improvements only if the ground-truth glomeruli are convex and the models systematically under-segment; but the paper does not verify either condition, and no failure analysis is provided. Without uncertainty quantification, the claim that CovHuSeg 'assures' better masks is not supported by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CovHuSeg, a post-processing method that takes a deep-learning model's predicted segmentation mask, extracts its contour, replaces that contour by its convex hull, and fills the hull to produce the final mask. The motivation is that kidney glomeruli are approximately ball-shaped, so enforcing convexity should remove holes and irregular protrusions. The method is evaluated on the KPIs2024 kidney pathology dataset with four models (UNet, UNet++, UNet3+, TransUNet), four training-data splits of varying size, and both normal and Gaussian-noised test sets. The authors report absolute Dice improvements between 0.005 and 0.033 across all configurations and conclude that CovHuSeg improves accuracy, especially in limited-data scenarios.","tokens_in":8616,"tokens_out":3872,"duration_ms":37353,"significance":"If the reported improvements are reproducible and robust, CovHuSeg would be a simple, architecture-agnostic, parameter-free post-processing step for kidney glomeruli segmentation, with no fitting and with immediate use of standard convex hull libraries. The paper also engages a relevant geometric prior that is plausible for glomeruli. However, the evidence as presented is not sufficient to establish the central claim: the tables report single-run Dice values with no variance, no repeated seeds, and no statistical tests, and the paper never evaluates the validity of the convexity assumption on the actual ground-truth shapes. These are fixable weaknesses, but they are load-bearing for the claim of universal improvement.","major_comments":[{"comment":"The central claim that CovHuSeg improves Dice for all four models, all four splits, and both normal and noisy test sets rests on a single training run per configuration. The reported absolute gains can be as small as 0.005 (Table 1, UNet Split B: 0.858 to 0.863) and 0.008 (Table 2, UNet Split B), which are on the order of typical run-to-run variation for segmentation networks trained for 150 epochs. The authors should provide multiple seeds (e.g., 3–5) with mean and standard deviation, plus per-image paired Dice differences and a significance test or confidence intervals. Without such uncertainty quantification, the claimed universal improvement cannot be distinguished from training stochasticity.","section":"Section 4.3, Tables 1 and 2"},{"comment":"The CovHuSeg operation is monotonic expansion: replacing the predicted mask by the filled convex hull of its contour adds pixels and removes none. It can therefore improve Dice only when the baseline systematically under-segments relative to a ground truth that is convex or nearly convex; it cannot repair over-segmentation or non-convex structures. The paper neither quantifies under- versus over-segmentation in the baseline predictions nor reports any case where the hull increases error. The favorable example in Figure 3 is not enough; the authors should add a failure analysis, such as a histogram of per-patch Dice change and examples of worsened patches, and an explicit assessment of the convexity of the ground-truth glomeruli.","section":"Section 3.3 and Figure 3"},{"comment":"The conclusion that CovHuSeg \"performs much better in scenarios with limited data\" is not supported by the reported tables. For UNet and UNet++, the absolute gains in Split D (25% of samples) are no larger than in Split A (50% of subjects, 100% of samples): for example, UNet normal gains are 0.005 for both Split B and Split D, while Split C gives 0.007. The larger gains for TransUNet and UNet3+ are confounded by their lower baseline Dice, and no statistical comparison across splits is provided. This claim should be removed or carefully qualified.","section":"Section 4.2 and Section 5"}],"minor_comments":[{"comment":"The introduction states that \"all models have increased accuracy in all metrics,\" but the experimental section reports only Dice. Either report additional metrics (e.g., IoU, precision, recall, boundary Dice) or revise the wording to say \"in the reported Dice metric.\"","section":"Section 1 and Section 4.3"},{"comment":"There are typos: \"Segmenation\" in the keywords and \"Correspondong author\" in the footnote should be corrected.","section":"Keywords and affiliation footnote"},{"comment":"The caption and surrounding text describe only a favorable example; the authors should state whether this patch is representative of the average behavior or a selected success case, since Figure 3 is otherwise easy to read as cherry-picking.","section":"Figure 3"},{"comment":"Reproducibility details are incomplete: the paper gives the optimizer, learning rate, and loss weights, but not batch size, input patch resolution, data augmentation, or the random seed used for the reported run. These details should be added.","section":"Section 4.2"},{"comment":"The sentence \"applying the CovHuSeg technique can improve the performance of these models compared to using the proposed method\" appears to contain a wording error; it should read \"compared to not using the proposed method.\"","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The core idea is simple and the direction is reasonable, but the empirical case is currently under-powered. The lack of multi-seed statistics is the main blocker; if the authors can supply repeated runs, paired per-image analyses, and an honest failure analysis, the paper could become acceptable. I would also ask the editor to confirm that the KPIs2024 challenge dataset is publicly available in a way that permits the reported evaluation to be reproduced."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper takes a known idea—convex hull post-processing, already used in brain tumor [37] and lung CT [42] segmentation—and applies it to kidney glomeruli. It does that clearly and honestly: the algorithm is simple (contour, hull, fill), the experiments cover four models, four data splits, and a noisy test set, and the reported Dice gains are consistently positive. For weak models the gains are larger (up to ~0.03 Dice), which is exactly where a shape prior should help. So the core empirical pattern is believable.\n\nWhat’s genuinely useful: the paper quantifies the effect of a cheap post-processing step on the KPIs2024 dataset, and it is careful to split by subjects and samples, which is more than many short papers do. It also cites the prior convex hull work rather than hiding it.\n\nThe soft spots are real but proportionate. The biggest is statistical: each configuration is a single training run, and the gains are 0.005–0.033 Dice. For 150 epochs of training, run-to-run variation can easily be that large. Without multiple seeds or any paired significance test, the claim that CovHuSeg 'assures' better masks is too strong. That should be fixed with multi-seed runs, or at least bootstrap CIs over test patches. Second, the abstract says 'Unlike other post-processing methods' but the cited literature [37,42] already uses convex hulls as post-processing; the novelty is the domain, not the mechanism. Third, the convexity premise—glomeruli are near-convex and models under-segment—is never checked. If a model over-segments beyond the convex envelope, hulling can only enlarge the error, and no failure cases are shown.\n\nThese are fixable. The paper is short, clearly written, and the idea is not wrong; it just needs more evidence and a more measured claim. I’d send it to a workshop or a short-paper track, and ask for error bars, a comparison against plain convex hull baselines, and an analysis of cases where Dice decreases. With those, it’s a citable engineering note.\n\nVerdict: worth a serious referee, but as a minor contribution. If I were editing, I’d send it out.","headline":"A simple, honestly-cited convex hull post-processing trick that shows consistent but statistically unsupported Dice gains; worth a lightweight referee, not a headline result.","tokens_in":9137,"tokens_out":2133,"would_cite":false,"duration_ms":19940,"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":"Convex-hull post-processing of predicted segmentation masks improves Dice score for every tested deep-learning model on kidney glomeruli images, with absolute gains of 0.005–0.033 across normal and noisy test sets.","keywords":["kidney pathology","glomeruli segmentation","convex hull","post-processing","deep learning segmentation","Dice score","whole slide images","medical image segmentation"],"falsifier":"Measure the area gap between each ground-truth glomerulus mask and its own convex hull on the KPIs2024 test set; if a large fraction of ground-truth masks are significantly non-convex, then uniformly positive Dice gains from CovHuSeg would be hard to reproduce, and the reported improvements would have to be explained by something other than the convexity prior.","tokens_in":8242,"feed_emoji":"🔬","tokens_out":8039,"duration_ms":66056,"temperature":0.7,"pith_summary":"The paper proposes CovHuSeg, a post-processing step that takes a segmentation model's predicted mask, traces its contour, replaces that contour by the smallest enclosing convex polygon, and fills the polygon back into a mask. The claim is that this purely geometric repair improves segmentation accuracy because glomeruli are ball-shaped and should never appear as hole-riddled, concave, or fragmented shapes. The authors test the step on four deep-learning segmentation models (UNet, UNet++, UNet3+, and TransUNet) under four data-reduction splits of the KPIs2024 kidney pathology challenge, on both clean and Gaussian-noised test images. They report that Dice score rises for every model and every split, with absolute gains from 0.005 to 0.033, and that the relative gains are largest for the weakest model and for noisy inputs.","feed_headline":"Convex hull post-processing lifts Dice for every segmentation model","feed_subtitle":"Simple hull-based repair improves kidney glomeruli segmentation on clean and noisy images, with Dice gains up to 0.033.","key_machinery":"The central object is the convex hull of the closed contour extracted from a model's binary mask: the smallest convex polygon that contains every point of that contour. The CovHuSeg pipeline is four steps: extract the contour, compute its convex hull, then fill the resulting hull contour to produce a new mask. The hull is what injects the prior that a glomerulus is a compact, convex, hole-free blob, and it removes internal holes and forbids unusual shapes that cannot be glomeruli. The algorithm is implementation-agnostic about how the hull is computed, citing Quickhull and other standard routines as interchangeable choices.","core_discovery":"The paper's central claim is that requiring the output mask to be a filled convex polygon is a sufficient geometric prior for glomerulus segmentation: after CovHuSeg, every tested model scores higher Dice than its raw prediction in all sixteen model–split combinations. The absolute Dice increase is 0.005 to 0.032 on normal test patches and 0.008 to 0.033 on noisy test patches; the relative increase is larger for the models with lower starting accuracy, reaching 10.40% for TransUNet on the noisiest comparison. The authors interpret this as evidence that the convex hull repairs the kinds of mistakes deep networks make on this task, internal holes, concave dents, and fragmented boundaries, rather than merely rescoring confident pixels.","pith_inferences":["A natural testable extension is to apply CovHuSeg to instance-level metrics and boundary-aware measures, since filling holes and straightening contours should also improve those scores; the paper reports only Dice.","The near-monotone relationship between lower baseline accuracy and larger gains suggests the hull acts mainly on fragmented predictions; if so, its benefit should shrink as baseline quality approaches the annotation ceiling, a prediction that could be checked on near-perfect models.","One could fold convexity into training by adding a convex-hull-aware loss term, which the authors propose as future work; the mechanism implies this would reduce the artifacts the post-processor currently fixes.","A reader could test whether the gains partly come from matching the annotation style rather than histologic truth: if challenge ground truths were drawn as smooth convex polygons, the hull moves predictions toward the labels' geometric style, so re-evaluating on pixel-exact manual traces would separate the two effects."],"forward_implications":["Any deep-learning segmentation model for PAS-stained kidney glomeruli can be given a Dice gain of roughly 0.5 to 3.3 percentage points at inference time by applying CovHuSeg, with no retraining or architecture change.","The method matters most when training data are scarce: relative gains on reduced-subject and reduced-sample splits are consistently larger than on the full-data split for all four models.","CovHuSeg also recovers part of the accuracy lost when test images are corrupted by Gaussian noise, so it behaves as a cheap inference-time regularizer for ball-shaped targets.","Because the hull step operates only on the output mask, the same recipe should transfer to other ball-shaped anomalies, such as cell nuclei or follicles, wherever the ground-truth boundaries are near-convex."],"supporting_citations":[{"why":"UNet is the strongest baseline and one of the four architectures whose masks are improved by CovHuSeg in every data split.","marker":"[35]"},{"why":"UNet++ is a second baseline architecture, and its nested-skip design makes it a distinct model in the comparison.","marker":"[43]"},{"why":"UNet3+ is a baseline in the comparison, showing larger relative gains than the other UNet variants.","marker":"[20]"},{"why":"TransUNet is the transformer-based baseline that starts with the lowest accuracy and receives the largest absolute Dice gains.","marker":"[10]"},{"why":"CircleNet supplies the prior that glomeruli are ball-shaped objects, the geometric premise CovHuSeg replaces with a convex hull.","marker":"[41]"},{"why":"Prior work applying a convex hull filter to brain-tumor segmentation is cited as evidence that convexity post-processing can improve accuracy.","marker":"[37]"},{"why":"Lung contour repair with an improved convex hull is cited as a related demonstration that hull-based repair improves accuracy and robustness.","marker":"[42]"},{"why":"Quickhull is one of the standard convex-hull routines the algorithm can use at the hull-computation step.","marker":"[3]"}],"fun_headline_variants":["Convex hull post-processing boosts kidney glomeruli Dice","Hull repair lifts Dice for all tested kidney segmentation models","Convex hull raises Dice up to 0.033 on kidney images","Hull post-processing fills holes, boosts kidney segmentation accuracy","Simple convex hull fix improves Dice across all kidney models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every glomerulus in these PAS-stained images is convex or near-convex, so replacing a predicted mask with its filled convex hull moves the mask toward the ground truth rather than inflating it beyond the true object.","fun_headline_variants_meta":{"raw":{"variants":["Convex hull post-processing boosts kidney glomeruli Dice","Hull repair lifts Dice for all tested kidney segmentation models","Convex hull raises Dice up to 0.033 on kidney images","Hull post-processing fills holes, boosts kidney segmentation accuracy","Simple convex hull fix improves Dice across all kidney models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000644,"raw_usage":{"total_tokens":2907,"prompt_tokens":841,"completion_tokens":2066,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":1984}},"tokens_in":457,"tokens_out":2066,"duration_ms":14470,"temperature":1.0,"reasoning_tokens":1984,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:45:53.110993+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the area gap between each ground-truth glomerulus mask and its own convex hull on the KPIs2024 test set; if a large fraction of ground-truth masks are significantly non-convex, then uniformly positive Dice gains from CovHuSeg would be hard to reproduce, and the reported improvements would have to be explained by something other than the convexity prior.","supporting_citations":[{"cited_title":"In: Medical image computing and computer-assisted 12 Huy et al","cited_arxiv_id":null,"evidence_quote":"UNet is the strongest baseline and one of the four architectures whose masks are improved by CovHuSeg in every data split."},{"cited_title":"In: ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP)","cited_arxiv_id":null,"evidence_quote":"UNet3+ is a baseline in the comparison, showing larger relative gains than the other UNet variants."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention– MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part IV 23","cited_arxiv_id":null,"evidence_quote":"CircleNet supplies the prior that glomeruli are ball-shaped objects, the geometric premise CovHuSeg replaces with a convex hull."},{"cited_title":"In: Mahmud, M., He, J., Vassanelli, S., van Zun- dert, A., Zhong, N","cited_arxiv_id":null,"evidence_quote":"Prior work applying a convex hull filter to brain-tumor segmentation is cited as evidence that convexity post-processing can improve accuracy."},{"cited_title":"Scientific Reports 14 (07 2024)","cited_arxiv_id":null,"evidence_quote":"Lung contour repair with an improved convex hull is cited as a related demonstration that hull-based repair improves accuracy and robustness."},{"cited_title":"ACM Transactions on Mathematical Software (TOMS)22(4), 469–483 (1996)","cited_arxiv_id":null,"evidence_quote":"Quickhull is one of the standard convex-hull routines the algorithm can use at the hull-computation step."}],"review_version":1}