{"id":"ae6d3efd-24a9-44b3-898c-fcff6fc5b9d7","arxiv_id":"2508.15208","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"DyMorph-B2I combines watershed, skeletonization, and morphology with per-image tuning to separate renal structures from binary masks, but its evaluation is circular because tuning uses the ground-truth object counts.","lead":"This paper presents a pipeline that converts tissue-level binary masks into individual object segmentations by combining watershed, skeletonization, and morphological operations with image-specific parameter tuning. It claims near-perfect accuracy on renal pathology images, but the evaluation is undermined because the tuning uses the same human object counts that serve as the test labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation is circular: hyperparameters are tuned to match human object counts on the same images used for MAPE, so DyMorph-B2I's near-perfect accuracy is by construction.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: hyperparameter tuning uses human object counts as the reference, and the same human counts are used to compute MAPE/PE. This is not a minor methodological omission but a direct circularity that invalidates the reported quantitative comparison. The paper presents no evidence that the method can achieve such accuracy without access to the test labels. The central claim of outperforming classical baselines is therefore unsupported. The concern is concrete and testable; a proper held-out evaluation would either confirm or refute the claim. Given the explicit statements in §2.3 and Fig. 3, the current manuscript should be rejected, though a revised version with appropriate validation could be reconsidered.","tokens_in":8165,"tokens_out":3141,"duration_ms":38531,"concrete_test":"Hold out a random 20% of the 50 ROIs per class. Tune all DyMorph-B2I hyperparameters on the remaining 80% using only their human object counts (or choose a fixed grid without any test counts), then evaluate MAPE/PE on the held-out ROIs. Repeat 5-fold. If the held-out MAPE is not close to 0.0016 (e.g., >0.05), the reported accuracy is due to using the test labels during tuning.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim that DyMorph-B2I outperforms classical baselines rests on the evaluation protocol in §2.3 and §3.2. Section 2.3 states that 'Hyperparameter tuning was guided by expert anatomical knowledge, with the object count serving as a primary reference for evaluating plausibility,' and Fig. 3 caption confirms 'Evaluation was based on the object number determined by clinical anatomy knowledge, i.e., the manual annotation shown in the figure.' MAPE/PE in §3.2 are computed against the same human object counts. Thus the optimization procedure has direct access to the test labels: for each image, the grid search selects the parameter tuple that makes the predicted count match the human count. With seven tunable parameters (method, dist_kernel, fg_thresh, min_len, max_len, min_area, inter_collect), matching a single scalar count is almost always possible. The baselines are not given this label-dependent tuning, so the comparison is unfair. The reported average MAPE of 0.0016 reflects fitting to the labels, not genuine segmentation ability. Without a held-out split or a clear separation between tuning and evaluation, the claim is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DyMorph-B2I, a pipeline that combines watershed, skeletonization, and morphological operations with image- and class-specific hyperparameter tuning to convert binary semantic masks into instance segmentations for renal pathology structures (DT, PT, CAP, TUFT, ART, PTC). The method is evaluated on 300 ROIs from the NEPTUNE project using mean absolute percentage error (MAPE) and percentage error (PE) of instance counts against human annotations. The authors report average MAPE of 0.0016 and Spearman/Pearson correlations above 0.99, and conclude that DyMorph-B2I outperforms individual classical methods and a naive combination.","tokens_in":8461,"tokens_out":2130,"duration_ms":25809,"significance":"If the reported evaluation were valid, DyMorph-B2I would be a useful, transparent post-processing tool for converting available binary masks into instance counts, and the public code release is valuable. However, the central evaluation is circular: hyperparameters are selected using the human object count as the reference on the same images that are later used to compute MAPE/PE. As a result, the near-perfect error values (e.g., 0.0000 for several classes in Table 1) reflect fitting to the evaluation labels rather than independent generalization. The paper's main contribution is therefore not currently supported by the evidence presented.","major_comments":[{"comment":"The evaluation procedure is circular. Section 2.3 states that hyperparameter tuning was 'guided by expert anatomical knowledge, with the object count serving as a primary reference for evaluating plausibility,' and Fig. 3 confirms that evaluation is based on the object number determined by manual annotation. The same human object counts are then used in Eq. (1) and Eq. (2) of Section 3.2 to compute MAPE and PE, and in Table 1 to report near-zero errors. The grid search over at least seven tunable parameters (method, dist_kernel, fg_thresh, min_len, max_len, min_area, inter_collect) has direct access to the test labels. Matching a single scalar count per image is therefore almost guaranteed, and the reported average MAPE of 0.0016 is minimized by construction rather than being an independent measure of segmentation accuracy. This invalidates the central claim in Section 4.1 that DyMorph-B","section":"§2.3, §3.2, Table 1, Fig. 3"},{"comment":"No held-out validation or data split is described. The paper reports per-class statistics over 50 ROIs per class, but it never states whether the hyperparameter selection in Section 2.3 was performed on a separate tuning set, a training fold, or the same images used for evaluation. Without a visible separation between the images used for tuning and those used for evaluation, the quantitative results in Table 1 cannot be interpreted as generalization performance. The authors should provide a clear split (e.g., per-image or per-ROI cross-validation) and report MAPE/PE on held-out images only, with the tuning procedure applied without access to the human counts for those images.","section":"§3.1–§3.2"},{"comment":"The comparison against baselines is unfair because DyMorph-B2I receives image-specific, label-guided tuning while the classical methods (FindContour, Watershed, Skeleton, Morphology, and the naive combination) are evaluated with fixed, presumably untuned parameters. Even if the tuning were done on a separate set, the comparison should be controlled: either tune all methods under the same protocol, or evaluate all methods with parameters selected only from a training subset. As written, the 'win' over baselines may be due entirely to access to the evaluation labels, not to the proposed integration and refinement steps.","section":"Table 1, §2.3"}],"minor_comments":[{"comment":"The caption states 'Positive values indicate under-segmentation by the method, while negative values indicate over-segmentation.' This appears reversed: with PE = (m_i - h_i)/h_i, a positive value means the method produced more instances than the human reference, i.e., over-segmentation, and a negative value means under-segmentation. Please correct.","section":"Fig. 5 caption"},{"comment":"There are several typos and formatting inconsistencies, e.g., 'W/O Tunning' in Fig. 4, 'naïve' vs. 'naive', and inconsistent use of 'Error' in Table 1 without stating that it is MAPE.","section":"Throughout"},{"comment":"Reference [17] appears to be a duplicate of reference [11]. Also, reference [22] has a typo in the author name ('be, K.' should probably be 'Abe, K.').","section":"References"},{"comment":"The description of 'mix params' lists 'five essential elements' and then enumerates five items (method, dist kernel, fg thresh, min len, max len), which is consistent, but the naming 'mix params' is confusing and should be defined more precisely.","section":"§2.3"},{"comment":"Fig. 5 shows boxplots of PE but no statistical tests, confidence intervals, or per-image variability are provided. Since the central comparison is affected by the circularity issue, this is only a presentation concern once the evaluation protocol is fixed.","section":"§4.1"}],"recommendation":"reject","confidential_remarks":"The paper has a public code repository and a clearly described pipeline, but the evaluation is circular in a load-bearing way. I do not see how the current claims can be salvaged without redoing the experiments under a proper train/validation/test protocol. The manuscript may be resubmitted after such a re-evaluation, but in its present form it does not meet the standard for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on DyMorph-B2I. The core idea is a classical image-processing pipeline—watershed, skeletonization, morphology, convex-defect-guided line refinement—with per-image hyperparameter search to separate touching instances in renal pathology binary masks. That integration is real and honestly described; the components are standard and cited. The paper also ships a public GitHub repo, which is good. So there is a legitimate engineering contribution here, modest in scale.\n\nThe problem is the evaluation. Section 2.3 states that tuning is 'guided by expert anatomical knowledge, with the object count serving as a primary reference.' Figure 3 says evaluation 'was based on the object number determined by clinical anatomy knowledge, i.e., the manual annotation shown in the figure.' Section 3.2 then computes MAPE and PE against exactly those human counts. So the grid search has direct access to the test labels. With seven tunable parameters, matching a single scalar count is trivial, and the near-zero errors in Table 1 (0.0000 for CAP, TUFT, ART, PTC) are exactly what you'd expect from fitting, not from genuine segmentation ability. The baselines don't get the same label-dependent tuning, so the comparison is unfair and the claim that DyMorph-B2I 'outperforms' them is unsupported.\n\nThere's no held-out split, no description of how the tuning would work without labels at inference time, and no comparison to any learning-based method. The paper's own limitations section acknowledges accuracy drops on hard classes, but the reported numbers don't reflect that.\n\nI don't think this is deliberate gaming; the authors are transparent about the tuning protocol. But transparency doesn't fix circularity. What's needed is a proper split: tune parameters on a validation set, then report test MAPE, or at least describe a parameter-free protocol that doesn't look at the test counts.\n\nThe paper is coherent and addresses a real problem, and the pipeline could be useful to pathologists if the evaluation is redone properly. As it stands, the title claim 'high accuracy' is not supported. For a journal, I'd send it back for major revision with a clear request for a non-circular evaluation. For a conference, it shouldn't be accepted in this form.","headline":"Useful engineering integration, but the evaluation is circular: parameters are tuned to match the same human counts used for scoring, so the near-perfect numbers are by construction.","tokens_in":8930,"tokens_out":2032,"would_cite":false,"duration_ms":21850,"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":"DyMorph-B2I claims that a dynamic, morphology-guided pipeline—combining watershed, skeletonization, and morphological operations with per-class hyperparameter tuning—converts binary renal-pathology masks into instance counts that match huma","keywords":["binary-to-instance segmentation","renal pathology","instance segmentation","watershed","skeletonization","mathematical morphology","morphometric analysis","hyperparameter tuning"],"falsifier":"Run a held-out evaluation: tune the five hyperparameters per class on half of the ROIs using only those images' human counts, freeze the parameters, and apply the pipeline to the held-out images with no human reference. If the held-out MAPE climbs from 0.0016 toward the tens of percent seen for the classical baselines, the reported superiority is a tuning artifact; if it stays near 0.0016, the transfer claim holds.","tokens_in":8096,"feed_emoji":"🔬","tokens_out":9595,"duration_ms":101170,"temperature":0.7,"pith_summary":"This paper tackles a practical bottleneck in computational renal pathology: semantic segmentation models produce binary masks that lump many touching structures together, so object counts and per-structure measurements are unreliable. DyMorph-B2I is a post-processing pipeline that converts those binary masks into separated instances by chaining three classical techniques—watershed, skeletonization, and morphological operations—and then refining the proposed cut lines with geometry-aware rules (convex-defect depth selection, skeleton alignment, length constraints) plus per-class hyperparameter tuning. On 300 annotated ROIs spanning six renal structure classes, the paper reports an average count error of 0.16% and correlations above 0.99 with human expert counts, far below the error of each classical method alone or the naive combination. The claim matters because it offers a transparent, retraining-free route from the binary masks most datasets already provide to the instance-level quantification that diagnosis and research actually need.","feed_headline":"Kidney binary masks to instance counts: 0.16% mean error","feed_subtitle":"Watershed, skeleton, morphology plus per-class tuning match human counts on six renal structures.","key_machinery":"The load-bearing mechanism is the geometry-guided refinement step that turns raw watershed and skeleton outputs into plausible cut lines. Convex defect depth selection identifies the two deepest concavities of a candidate region and uses them as provisional endpoints; those endpoints are then compared with the region's skeleton, an approximation of the true medial axis, and shifted to align with it, so cuts follow anatomically plausible boundaries rather than arbitrary chords. A bisection-based length constraint rejects cut lines that are too short (noise) or too long (over-segmentation), and the whole configuration is searched per class over a defined hyperparameter tuple—method, distance-t","core_discovery":"The paper claims that binary-to-instance conversion in renal pathology is not solved by any single classical operator, but can be solved by their combination under morphology-guided, image-adaptive control. In the proposed pipeline, watershed first separates roughly circular adhesions; convexity-defect analysis then finds concave junctions that are likely true boundaries, using the two deepest defects as provisional endpoints for cut lines that are aligned to the region skeleton to approximate the medial axis; morphological operations and length constraints clean spurious cuts; and a five-parameter search (segmentation method, distance-kernel size, foreground threshold, minimum and maximum c","pith_inferences":["A fair reading of Section 2.3 suggests the 0.0016 MAPE is an optimistic bound: the tuning reference and the evaluation target are the same human counts, so a deployment protocol with parameters frozen after tuning on a labeled subset would likely show a larger—though not necessarily unattractive—error.","The dataset spans four stains (H&E, PAS, silver, trichrome) but results are pooled; whether per-class parameters transfer across stains or must be re-derived per stain is open and directly testable from the released pipeline.","The metrics count objects only; if the pipeline's real value is morphometric, per-instance boundary fidelity (for example, Dice or boundary distance) must be measured, since counting can be correct while geometry is wrong.","Deterministic and invertible in its parameters, the pipeline could serve as a pseudo-label engine to distill morphology-guided priors into a fast network, an extension of the paper's stated direction that would also make computational cost acceptable at scale."],"forward_implications":["Any binary semantic mask of renal structures, whatever segmentation model produced it, can be upgraded to instance-level counts and geometry by a deterministic pipeline with no training—only per-class parameter values.","The same code path adapts to the morphology of each structure class, so quasi-circular objects (capsules, tufts) and elongated or densely packed ones (tubules, arteries, peritubular capillaries) are handled by the same mechanism.","Because count errors approach zero across all six classes, downstream morphometric quantities such as size, shape, and spatial distribution become computable from existing binary-mask datasets.","The pipeline is fully transparent, combining classical operators with explicit, inspectable parameters, which makes it auditable in settings where black-box components are unacceptable.","The method can generate instance pseudo-labels for training learning-based instance segmenters, converting a slow classical post-process into cheap supervision, as the paper's future-work section anticipates."],"supporting_citations":[{"why":"Supplies the watershed transform that performs the initial separation of roughly circular adhesions in the binary mask.","marker":"[18]"},{"why":"Supplies the morphological erosion and dilation operators used to introduce separation points and clean boundaries.","marker":"[19]"},{"why":"Supplies the thinning algorithm that yields the region skeleton used as the medial-axis reference for aligning candidate cut lines; the paper's in-text attribution differs from its reference-list author.","marker":"[20]"},{"why":"Source of the 300 ROI images and the manual expert counts that the evaluation treats as ground truth.","marker":"[21]"},{"why":"The contour-extraction baseline (FindContour) that the method must beat; its border-following principle also underlies the minimum-area contour filtering in the final stage.","marker":"[22]"}],"fun_headline_variants":["Kidney instance segmentation: 0.16% mean error via hybrid method","Watershed + skeleton + morphology: 0.16% kidney mask error","DyMorph-B2I: dynamic morphology-guided splitting for kidney pathology","Combining classical ops beats single methods in kidney instance counts","Per-class tuning lifts kidney binary-to-instance to 0.16% error"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the per-class hyperparameters tuned against human object counts (Section 2.3) will transfer to new images without access to those counts, yet the paper reports no held-out split, leaving open the possibility that the near-zero errors reflect fitting to the evaluation labels rather than genuine segmentation ability.","fun_headline_variants_meta":{"raw":{"variants":["Kidney instance segmentation: 0.16% mean error via hybrid method","Watershed + skeleton + morphology: 0.16% kidney mask error","DyMorph-B2I: dynamic morphology-guided splitting for kidney pathology","Combining classical ops beats single methods in kidney instance counts","Per-class tuning lifts kidney binary-to-instance to 0.16% error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000587,"raw_usage":{"total_tokens":2578,"prompt_tokens":712,"completion_tokens":1866,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":1778}},"tokens_in":456,"tokens_out":1866,"duration_ms":12694,"temperature":1.0,"reasoning_tokens":1778,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:01:11.786912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a held-out evaluation: tune the five hyperparameters per class on half of the ROIs using only those images' human counts, freeze the parameters, and apply the pipeline to the held-out images with no human reference. If the held-out MAPE climbs from 0.0016 toward the tens of percent seen for the classical baselines, the reported superiority is a tuning artifact; if it stays near 0.0016, the transfer claim holds.","supporting_citations":[{"cited_title":"An efficient immersion-based watershed transform method and its prototype ar- chitecture,","cited_arxiv_id":null,"evidence_quote":"Supplies the watershed transform that performs the initial separation of roughly circular adhesions in the binary mask."},{"cited_title":"Image analysis using mathematical morphology,","cited_arxiv_id":null,"evidence_quote":"Supplies the morphological erosion and dilation operators used to introduce separation points and clean boundaries."},{"cited_title":"A fast parallel algorithm for thinning digital patterns,","cited_arxiv_id":null,"evidence_quote":"Supplies the thinning algorithm that yields the region skeleton used as the medial-axis reference for aligning candidate cut lines; the paper's in-text attribution differs from its reference-list author."},{"cited_title":"Digital pathology evaluation in the multicenter nephrotic syndrome study network (neptune),","cited_arxiv_id":null,"evidence_quote":"Source of the 300 ROI images and the manual expert counts that the evaluation treats as ground truth."},{"cited_title":"Topological structural analysis of digitized binary images by border following,","cited_arxiv_id":null,"evidence_quote":"The contour-extraction baseline (FindContour) that the method must beat; its border-following principle also underlies the minimum-area contour filtering in the final stage."}],"review_version":1}