{"id":"d78aad1a-a076-4d19-a9ff-f8a03290de93","arxiv_id":"2505.05073","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RepSNet combines four-direction boundary distance regression with a boundary voting mechanism and reparameterizable encoder-decoder to reach mPQ 0.5633 on the authors' Lizard split and 0.478 on the official CoNIC test set.","lead":"A new deep learning model, RepSNet, separates overlapping cell nuclei in stained pathology images by having each nucleus pixel estimate the distance to its nucleus wall in four directions, and then letting those pixels vote on where the wall actually sits. On the standard Lizard dataset it reports higher accuracy and faster training convergence than several published baselines, and it took second place on the public CoNIC challenge leaderboard.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed mPQ gain over StarDist is measured on a random patch split of Lizard (Sec. 4.1), not the official CoNIC split; the paper's own official result (Sec. 4.6, Table 4) shows NBRIA behind StarDist by 0.023, so the 'outperforms SOTA' claim is unsupported.","rationale":"The reader's weakest assumption focuses on the non-differentiability of the boundary voting loss Lnb in Eq. (6). That is a genuine internal consistency issue, but it concerns one component of the method and can be settled by code inspection; the official second-place leaderboard result would survive even if Lnb is ineffective. The concern I identify is more directly load-bearing for the central claim as stated in the reader's strongest_claim: the assertion that NBRIA 'significantly outperforms other benchmark models' and 'improves SOTA StarDist by 0.0161' is based on a non-official random patch split in Section 4.1, while the paper's own official evaluation in Section 4.6 shows the opposite ranking. The validity of the random split is therefore the weakest link in the headline empirical claim. I set verdict_should_be to UNCHANGED because the paper still provides a credible second-place official result, so a conditional verdict requiring re-evaluation on an independent split and removal of the unsupported SOTA claim is appropriate, matching the reader's condition. Agreement is partial because the reader identified the Lnb gradient as weakest but did mention leakage risk in the rationale; my concern aligns with that secondary point rather than the reader's chosen assumption.","tokens_in":24569,"tokens_out":10360,"duration_ms":111850,"concrete_test":"Use the released dataset splitting configuration (github.com/luckyrz0/RepSNet) together with Lizard/CoNIC metadata to assign each patch its source whole-slide image ID. If any patch in the local test split shares a source slide with a training patch, the split is not slide-exclusive. Then recompute Table 1 on the official CoNIC split (or on a slide-exclusive split) and compare mPQ against Table 4. If the mPQ gap versus StarDist reverses or drops below the reported 0.0161, the claimed outperformance of the CoNIC SOTA model is an artifact of the random patch-level split.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that NBRIA(RepSNet) 'significantly outperforms' benchmark models, including the CoNIC SOTA StarDist, by 0.0161 in mPQ (Abstract and Section 4.3.1, Table 1). This number comes from the locally constructed test set described in Section 4.1, where the 4,981 Lizard patches are 'randomly divided into training, validation, and test sets in a ratio of 7:1:2.' Lizard patches are extracted from whole-slide images, and a random patch-level split can place patches from the same source slide in both training and test sets. This creates a leakage path: the model can memorize slide-specific staining, artifacts, and tissue texture, inflating test metrics relative to a slide-exclusive or official split. The paper itself provides evidence that this inflation is large: Table 4, the official CoNIC final test set, lists NBRIA(RepSNet) at mPQ 0.478 and EPFL(StarDist) at 0.501, i.e., the proposed method is behind StarDist by 0.023. Thus the contribution bullet 'Compared with the CoNIC SOTA StarDist model, improves it by 0.0161' is contradicted by the official, externally evaluated leaderboard. The load-bearing assumption, that the random patch split gives a valid comparison for the headline SOTA claim, is therefore insecure. The official second-place result is real evidence of near-SOTA performance, but the paper's primary framing of outperforming StarDist rests on the potentially leaked local split.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a nucleus instance segmentation and classification scheme, NBRIA(RepSNet), based on boundary distance regression, a boundary voting mechanism (BVM), and a structurally re-parameterizable encoder-decoder network with RepVGG/RepUpsample modules and a boundary isoheight loss. The method is evaluated on the Lizard/CoNIC dataset against U-Net, DCAN, Hover-Net, and StarDist, reporting an mPQ of 0.5633 on a locally constructed test split and 0.478 on the official CoNIC final test set, where it ranks second behind StarDist.","tokens_in":24986,"tokens_out":4084,"duration_ms":41728,"significance":"The boundary-voting idea and the use of structural re-parameterization for an encoder-decoder segmentation network are interesting and potentially useful. The public release of code, dataset split configuration, and pretrained models is a concrete strength, and the official CoNIC second-place result provides independent, externally evaluated evidence that the method is near the state of the art. However, the paper's primary claim that NBRIA(RepSNet) outperforms the CoNIC SOTA StarDist is contradicted by the paper's own official leaderboard result, and the boundary loss as formulated has no defined gradient through the voting procedure. These issues are load-bearing for the paper's central claims and must be resolved before the paper can be accepted.","major_comments":[{"comment":"The headline claim that NBRIA(RepSNet) 'significantly outperforms' StarDist by 0.0161 mPQ rests entirely on the locally constructed random patch split described in Section 4.1. Because Lizard patches are extracted from whole-slide images, a random patch-level 7:1:2 split can place patches from the same source slide in both training and test sets, allowing the model to memorize slide-specific staining and tissue texture. The paper itself provides strong evidence of this inflation: the official CoNIC final test set (Table 4) lists NBRIA(RepSNet) at mPQ 0.478, behind EPFL(StarDist) at 0.501. The abstract and contribution bullet in Section 1 therefore overstate the result. The authors should either re-evaluate all comparisons on a slide-exclusive split or present the official leaderboard as the primary basis for SOTA claims.","section":"Section 4.1, Table 1, Table 4"},{"comment":"The boundary loss Lnb in Eq. (6) is computed over the set NB, which is produced by Algorithm 1 via the hard thresholding operation 'NB > e_t' applied to the voting counts. This operation has no defined gradient with respect to the network parameters, yet Eq. (1) states that the total loss 'jointly optimizes' NP, NT, BD, and NB. As written, the gradient of Lnb cannot reach the weights of RepSNet through the voting step. If the authors intend a straight-through estimator, a soft approximation, or a detached NB, they must state this explicitly; otherwise the ablation results in Tables 2 and 3 cannot be attributed to Lnb, because the loss would be constant during training.","section":"Section 3.2.3, Eq. (6), Algorithm 1"},{"comment":"The boundary voting threshold e_t is selected 'based on mPQ metric' on the same locally constructed evaluation split used to report the headline mPQ of 0.5633. If this selection was performed on the test split, the reported numbers are optimistically biased. Please specify which split (validation vs. test) was used for threshold selection. If the test split was used, the comparison in Table 1 should be repeated with the threshold fixed before any test-set evaluation.","section":"Section 4.5.1"}],"minor_comments":[{"comment":"The word 'repectively' should be 'respectively'.","section":"Fig. 1 caption"},{"comment":"FLOPs(B) is used to denote floating-point operations, but the text says 'floating-point operations per second'; the former is a count and the latter is a rate. Please use separate notation for FLOPs and FPS.","section":"Section 4.2"},{"comment":"The notation 'PQ +' is not defined; please clarify whether it denotes per-class PQ, PQ^+ as used in the CoNIC challenge, or something else.","section":"Section 4.6, Table 5"},{"comment":"The phrase 'thoroughgoing re-parameterization model' is awkward; consider 'fully re-parameterizable model'.","section":"Section 3.2.1"},{"comment":"The sentence 'For Hover-Net, in order to enable the model to learn horizontal-vertical distance maps, the authors used ResNet-101 as its backbone network' is a comparison rationale but could be better supported by citing the Hover-Net paper's exact architecture details; currently it reads as an unsupported claim.","section":"Section 4.3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a credible core contribution and a real second-place result on the official CoNIC leaderboard, but the framing overstates the SOTA claim, and the boundary loss formulation as written is not trainable. The authors should be asked to fix the gradient issue, clarify threshold selection, and reposition the SOTA comparison around the official split. If these are addressed, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nRead RepSNet. Short version: it's a genuine methods paper with a novel boundary-voting pipeline and an official second-place result on CoNIC, but the headline claim that it beats StarDist is not supported by the paper's own numbers. The official leaderboard (Table 4) puts NBRIA at 0.478 mPQ versus StarDist at 0.501, a 0.023 gap. The 0.5633 mPQ in the abstract comes from a random patch-level split of Lizard (Section 4.1), where patches from the same whole slide can land in both training and test sets. That is a leakage path, and it fits the large local-to-official drop (0.563 to 0.478) for the same model. The 'outperforms SOTA' claim should be dropped or reframed as 'competitive on the official leaderboard.'\n\nCredit where due: regressing per-pixel distances to the parent nucleus boundary in four directions and then fusing them with a voting mechanism is a real alternative to centroid-based methods like Hover-Net and StarDist. RepUpsample is a reasonable extension of RepVGG to deconvolution, and the isoheight boundary loss is a sensible way to penalize boundary deviations without expensive distance transforms. The ablations are thorough, the code and pretrained models are shared, and the official second-place result is independent evidence that the method works.\n\nSoft spots, in rough order of severity. (1) The random patch split invalidates the main comparison, and Table 4 undercuts the abstract. (2) The boundary loss Lnb in Eq. (6) is computed on NB, which comes from a hard threshold on votes (Algorithm 1). There is no defined gradient through that step, yet the paper says the full loss jointly optimizes NB. If they use a straight-through or soft approximation, it needs to be stated; as written, the gradient path is missing. This is a real technical gap, not a nitpick. (3) The voting threshold was tuned on the same local split used for the headline numbers, so the 0.5633 has a selection effect. (4) No error bars or multiple seeds, which matters for the word 'significantly.'\n\nOverall: a competent, honest paper with a useful method and an overclaimed headline. It deserves peer review, but revision should fix the split, the gradient explanation, and the claims. I would not cite the 0.5633 number, but I would cite the method and the official leaderboard result.","headline":"A real method with a solid official benchmark result, undermined by a leaky local split and an unexplained boundary-loss gradient.","tokens_in":25484,"tokens_out":2858,"would_cite":true,"duration_ms":28221,"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":"The paper claims that letting every pixel of a nucleus vote on its parent boundary, then cutting instances by connected components, beats pixel-by-pixel boundary classification on the Lizard benchmark, reaching mPQ 0.5633.","keywords":["nucleus instance segmentation","boundary regression","boundary voting mechanism","structural re-parameterization","RepSNet","histopathological images","Lizard dataset"],"falsifier":"In the released training code, trace whether gradients from $L_{\\mathrm{nb}}$ reach the encoder and decoders through the voting step; then retrain with the $L_{\\mathrm{nb}}$ term removed, keeping BVM at inference, and check whether the mPQ gap in the paper's ablation persists across several seeds.","tokens_in":1580,"feed_emoji":"🔬","tokens_out":1973,"duration_ms":81471,"temperature":0.7,"pith_summary":"The paper proposes a scheme, NBRIA, in which nucleus instance segmentation is treated as a boundary-regression problem rather than a pixel-classification problem. For every pixel inside a nucleus, the network predicts the distance to that nucleus's boundary in four directions; a voting mechanism then accumulates those predictions from all pixels of the nucleus and keeps only boundary positions receiving enough votes. On the Lizard dataset the resulting model, RepSNet, reports mPQ 0.5633, above the StarDist baseline at 0.5472, with an inference rate of 10 images of 256x256 pixels per second. The authors also report second place on the CoNIC official test leaderboard (mPQ 0.478) and note that the online system closed before an offline ensemble version could be submitted.","feed_headline":"Boundary voting lifts nucleus segmentation to mPQ 0.5633","feed_subtitle":"Fusing boundary votes from many pixels beats StarDist and Hover-Net on Lizard at 10 images per second.","key_machinery":"The Boundary Voting Mechanism (BVM) is the load-bearing object: it aggregates boundary position information (BPI) from every foreground pixel's four-direction distance prediction, maps those distances to absolute coordinates, counts votes, and applies a hard threshold ($e_t=3$) to output a nucleus boundary map. This voting step is what converts a regression output into instance masks, and the paper argues it provides synergistic belief enhancement, so consistent small-error estimates reinforce each other while scattered large-error estimates cancel as noise. The two supporting pieces are the structural re-parameterization (RepVGG units plus RepUpsample deconvolution branches) that gives training-time multi-scale branches and inference-time single-branch efficiency, and the boundary isoheight loss $L_{\\mathrm{nb}}$ that penalizes boundary deviations adaptively.","core_discovery":"The central claim is that a nucleus boundary is more reliably located by fusing weak per-pixel boundary-distance estimates than by classifying boundary status directly. RepSNet computes a boundary-distance (BD) map with four channels that give, for each nucleus pixel, the distances to its parent nucleus boundary up, down, left, and right. The boundary voting mechanism converts these relative distances into absolute boundary positions, counts how many pixels vote for each position, and keeps positions whose vote count exceeds a threshold ($e_t=3$). Connected-component analysis on the thresholded boundary map yields instances. The boundary loss $L_{\\mathrm{nb}}$ uses morphological isoheights around the annotated boundary to penalize predicted boundary pixels by their approximate Chebyshev distance from the annotation, which the authors argue resists annotation noise. A fully re-parameterizable encoder-decoder, built from RepVGG units and a proposed RepUpsample module, gives multi-scale receptive fields during training and collapses to single-branch convolutions at inference.","pith_inferences":["Editorial inference: the BVM recipe suggests any distance-regression nucleus model, not only RepSNet, could replace its watershed or NMS post-processing with voting and gain robustness on overlapping nuclei; the paper's own exterior ablations on Hover-Net and StarDist already point in this direction.","Editorial inference: the paper's remark about adding more distance directions implies a testable scaling law, where increasing from four to eight or radial directions should improve recall on highly irregular nuclei but will raise the BD map's channel count and data demand.","Editorial inference: because the voting threshold is a hard step inside the loss path, a reader auditing the released code should check whether $L_{\\mathrm{nb}}$ reaches the weights through a straight-through or softened approximation; if it is detached, the reported gain of $L_{\\mathrm{nb}}$ would need a different explanation."],"forward_implications":["On the Lizard test split, the reported mPQ of 0.5633 exceeds StarDist (0.5472) and Hover-Net (0.5314), with higher AJI, DICE, and PQ among the compared methods.","The model processes about 10 images of 256x256 pixels per second with 18.6M parameters, which is faster than Hover-Net and StarDist in the paper's measurements.","Replacing the post-processing of Hover-Net and StarDist with BVM improves their AJI, DICE, PQ, and mPQ, so the voting mechanism transfers to other distance-based nucleus models.","Adding the boundary isoheight loss $L_{\\mathrm{nb}}$ to StarDist and DCAN improves their metrics, indicating the loss transfers beyond RepSNet.","Under equal training budgets, RepSNet reaches higher mPQ than StarDist at the same epoch count and at a fixed five-hour training time."],"supporting_citations":[{"why":"Supplies the Lizard dataset, the benchmark whose test split all reported mPQ comparisons use.","marker":"[13]"},{"why":"Defines Hover-Net, a primary NCDIP baseline, and supplies the DICE, AJI, and PQ metric conventions used in evaluation.","marker":"[10]"},{"why":"StarDist is the CoNIC state-of-the-art baseline that NBRIA(RepSNet) is claimed to beat on mPQ.","marker":"[26]"},{"why":"Supplies the RepVGG re-parameterizable unit whose multi-branch training and single-branch inference design is the core of RepSNet's efficiency.","marker":"[12]"},{"why":"DCAN is the contour-aware pixel-discrimination baseline used to argue for voting-based boundary aggregation.","marker":"[6]"},{"why":"U-Net is the shared encoder-decoder structure that RepSNet extends and one of the compared baselines.","marker":"[22]"},{"why":"Define the CoNIC challenge, its official data split, and the leaderboard where NBRIA(RepSNet) reports second place.","marker":"[14, 15]"}],"fun_headline_variants":["Boundary voting boosts nucleus segmentation accuracy","RepSNet fuses boundary votes for faster nucleus segmentation","Efficient nucleus segmentation via boundary regression and voting","Weak boundary votes fuse into stronger nucleus instances"],"cache_read_input_tokens":27520,"weakest_assumption_plain":"The training loop assumes that the boundary loss $L_{\\mathrm{nb}}$ can steer the network weights even though the boundary map it penalizes is produced by counting votes and applying a hard threshold, an operation with no defined gradient.","fun_headline_variants_meta":{"raw":{"variants":["Boundary voting boosts nucleus segmentation accuracy","RepSNet fuses boundary votes for faster nucleus segmentation","Efficient nucleus segmentation via boundary regression and voting","Weak boundary votes fuse into stronger nucleus instances"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"total_tokens":4231,"prompt_tokens":1053,"completion_tokens":3178,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":3120}},"tokens_in":669,"tokens_out":3178,"duration_ms":21633,"temperature":1.0,"reasoning_tokens":3120,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:13:53.529816+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the released training code, trace whether gradients from $L_{\\mathrm{nb}}$ reach the encoder and decoders through the voting step; then retrain with the $L_{\\mathrm{nb}}$ term removed, keeping BVM at inference, and check whether the mPQ gap in the paper's ablation persists across several seeds.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV) Workshops, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the Lizard dataset, the benchmark whose test split all reported mPQ comparisons use."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11, pp","cited_arxiv_id":null,"evidence_quote":"StarDist is the CoNIC state-of-the-art baseline that NBRIA(RepSNet) is claimed to beat on mPQ."},{"cited_title":"Medical Image Anal- ysis 36, 135–146 (2017)","cited_arxiv_id":null,"evidence_quote":"DCAN is the contour-aware pixel-discrimination baseline used to argue for voting-based boundary aggregation."}],"review_version":1}