{"id":"048766c3-3cc6-4682-ba0e-0a4d5e5f5258","arxiv_id":"2411.17557","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"BR-Net, a Mask R-CNN extension with bilayer overlap decomposition and semantic consistency regularization, reports higher segmentation accuracy than comparison methods on two private C. elegans datasets.","lead":"This paper proposes a neural network called BR-Net that separates overlapping roundworms in microscope images better than existing segmentation methods. It matters because accurate worm segmentation helps automate aging and drug studies in C. elegans, a standard lab organism.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed superiority is not verifiable because baseline methods lack a stated common training/evaluation protocol, and Table 1 citations conflict with the reference list.","rationale":"The reader's verdict is CONDITIONAL with the weakest assumption being the accuracy and consistency of the private annotations and derived sub-masks. I agree that label quality matters, but the more load-bearing issue is the lack of a controlled comparison protocol. The paper's central claim is explicitly comparative: BR-Net 'outperforms other recently proposed instance segmentation methods.' Without a common training/evaluation protocol, even perfectly accurate labels cannot substantiate the ranking. The internal citation mismatches (Mask R-CNN labeled [12], YOLACT labeled [13], etc.) are observable in the manuscript and directly affect the credibility of Table 1. This is an internal inconsistency, not a disagreement with consensus, and it is testable by re-running the baselines. The reader's conditionality already covers requiring artifact release and validation, so the verdict stays CONDITIONAL; my concern strengthens the justification for that condition without moving to a different verdict. I set agreement_with_reader to 'partial' because the reader identified label noise as the weakest assumption, whereas I identify the uncontrolled and mis-cited baseline comparison as the primary threat to the central claim.","tokens_in":12465,"tokens_out":4627,"duration_ms":43357,"concrete_test":"Re-run every method in Table 1 (Mask R-CNN, YOLACT, Mask Scoring R-CNN, Zeng et al., Li et al., Occlusion R-CNN, BCNet, Xiao et al., WormSwin, BR-Net) in a single framework (e.g., Detectron2/mmdetection) using identical training settings: same train/test split of C.Data-1 and C.Data-2, same ResNet-50-FPN backbone, same augmentation pipeline, same optimizer (SGD or Adam with fixed schedule, e.g., 12k iterations with warmup), and the same evaluation script computing AP, AP50, AP75, mIoU. If BR-Net still achieves the highest scores under this controlled comparison, the claimed superiority is supported; otherwise the Table 1 ranking is an artifact of unequal training conditions. Additionally, audit the reference numbering so that each named method is linked to its correct bibliography entry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BR-Net outperforms all comparison methods on C.Data-1 and C.Data-2 (Table 1: AP 0.739 vs 0.721 on C.Data-1; 0.712 vs 0.682 on C.Data-2). For this claim to hold, all methods must be trained and evaluated under identical, documented conditions. The paper never specifies the training setup for the baselines: for each baseline, there is no statement of which backbone, data augmentation, optimizer, learning rate schedule, or number of iterations was used. Section 4.3 only describes BR-Net's own setting (ResNet-50 FPN, batch size 8, initial LR 0.01, '200 iterations' with Adam — itself ambiguous, likely a typo for epochs). Without a shared protocol, the reported margins could stem from unfair hyperparameter tuning, longer training, or weaker baselines rather than architectural merit. The comparison is further undermined by internal citation inconsistencies: Table 1 labels '[12]' as Mask R-CNN, but reference [12] in the bibliography is SegNeXt (Guo et al., NeurIPS 2022); it labels '[13]' as YOLACT, but reference [13] is Zeng et al. on C. elegans segmentation; and it labels '[14]' as Zeng et al., but reference [14] is Wang et al. (Celeganser). These mismatches suggest the experimental table was not carefully checked against the cited literature, casting doubt on the provenance of the baseline numbers. The reader's concern about label quality is valid, but even with perfect labels, the ranking in Table 1 is not trustworthy unless the baselines are re-run under controlled conditions on the same private data with the same split and protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes BR-Net, a bilayer segmentation-recombination network for instance segmentation of overlapping C. elegans in microscope images. The architecture consists of a Coarse Mask Segmentation Module (CMSM, based on Mask R-CNN with a United/Unified Attention Module), a Bilayer Segmentation Module (BSM) that predicts overlapping and non-overlapping sub-regions, and a Semantic Consistency Recombination Module (SCRM) that merges sub-region predictions and enforces semantic consistency via a loss term. The method is evaluated on two privately collected C. elegans datasets, C.Data-1 and C.Data-2, against classic instance segmentation methods, amodal methods, and the worm-specific WormSwin. The authors report the highest AP, AP50, AP75, and mIoU on both datasets, with ablations in Tables 2 and 3 supporting the contribution of each module.","tokens_in":12832,"tokens_out":3516,"duration_ms":31460,"significance":"If the reported results are reproducible, the work is a useful and fairly well-motivated application of amodal-instance-segmentation ideas to an important biological imaging task. The decomposition into overlapping/non-overlapping regions followed by recombination is a sensible adaptation of recent amodal methods, and the structured ablation study gives insight into which components matter. However, the contribution is incremental with respect to existing amodal segmentation frameworks, the dataset is private, and no code or trained models are released. The paper does not provide machine-checked proofs or reproducible artifacts; its falsifiable content consists solely of performance numbers on a private benchmark. The significance is therefore conditional on the experimental comparison being made verifiable and statistically sound.","major_comments":[{"comment":"The training protocol for the baseline methods is not specified. Section 4.3 describes only BR-Net's own setting (batch size 8, initial learning rate 0.01, linear warmup in the first 20 iterations, 200 iterations with Adam). It is not stated which backbone, data augmentation, learning-rate schedule, or number of iterations was used for Mask R-CNN, YOLACT, BCNet, WormSwin, or any other baseline. Without a unified, documented protocol, the margins in Table 1 (e.g., AP 0.739 vs. 0.721 on C.Data-1) cannot be attributed to the architecture rather than to differences in training effort or hyperparameter tuning. Additionally, the phrase 'trained for 200 iterations' is ambiguous and likely a typo for epochs; as written, it would be fewer than one epoch on the 880-image training set, which is not a credible convergence setting for any of the methods.","section":"Section 4.3, Table 1"},{"comment":"The citation labels in Table 1 do not match the reference list. Table 1 cites [12] as Mask R-CNN, but reference [12] is SegNeXt (Guo et al., NeurIPS 2022). It cites [13] as YOLACT, but reference [13] is Zeng et al. on C. elegans segmentation. It cites [14] as Zeng et al., but reference [14] is Wang et al. (Celeganser). These mismatches make the provenance of the baseline numbers impossible to verify and suggest that the experimental table was not checked against the bibliography. The table must be rebuilt with correct citations or the references renumbered consistently.","section":"Table 1 and Reference List"},{"comment":"All quantitative results are reported as single numbers without error bars, number of repeated runs, seeds, or statistical significance tests. The claimed improvements over the strongest baseline are small (AP 0.739 vs. 0.721 on C.Data-1; AP75 0.760 vs. 0.737), and in the absence of variance estimates the superiority claim is not statistically supported. At minimum, the authors should report the mean and standard deviation over at least three independent runs, or provide a statistical significance test, for the main comparisons and ablations.","section":"Tables 1, 2, 3"},{"comment":"The overlapping and non-overlapping sub-mask labels are described as derived 'through logical operations based on the positional relationship between nematodes' (Section 3.1), but the concrete derivation rule is not given, and no validation of the resulting labels is presented. The dataset is private and the code is not released, so the central experiments cannot be independently reproduced. The authors should either release the data and code, or provide a detailed annotation protocol, a graphical example of the label derivation, and a quantitative assessment of label consistency, especially for heavily occluded instances where the decomposition is ambiguous and could bias the BSM and SCRM supervision.","section":"Sections 3.1 and 4.1"}],"minor_comments":[{"comment":"The module is called 'United Attention Module' in some headings and in Figure 4, but 'Unified Attention Module' in the abstract and elsewhere; the terminology should be made consistent.","section":"Throughout"},{"comment":"The sentence 'The network is trained for 200 iterations using Adam as the optimizer' should specify whether 'iterations' means epochs, and the total number of training steps should be stated.","section":"Section 4.3"},{"comment":"Equation (2) sums over images and instances but defines y_{i,c} without the image index k, and Equation (4) suppresses both indices; the notation should be consistent throughout.","section":"Equations (2)-(4)"},{"comment":"The caption contains a Chinese comma and the phrase 'utilises' with non-standard punctuation; this should be polished for typographic consistency.","section":"Figure 2 caption"},{"comment":"The text in Section 2.2 says 'In 2002, Sun et al.' but reference [21] is from 2022; the year is likely a typo.","section":"Reference [21]"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not acceptable in its current form because the experimental comparison, which is the core evidence for the claimed superiority, is not verifiable: baseline training protocols are missing, the citation labels in Table 1 conflict with the reference list, no statistical validation is provided, and the data and code are private. The architectural ideas are reasonable, and the work could become publishable after a major revision that reports a unified evaluation protocol, repeated runs with variance, corrected references, and a clear description (or release) of the dataset and sub-mask label derivation. I would also recommend that the editor ask the authors to clarify the '200 iterations' setting, as it currently appears to be a typo that further undermines confidence in the experimental report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: BR-Net is a plausible incremental architecture, but the paper's headline claim—that it beats every comparison method on both C. elegans datasets—cannot be evaluated from the manuscript because the baselines are not trained under a stated common protocol and the table's citations don't line up with the bibliography. I wouldn't take the AP numbers at face value, but I also wouldn't dismiss the method.\n\nWhat's genuinely new: the paper repackages three known ideas (Mask R-CNN, CBAM-like attention, BCNet's bilayer decomposition) into a single pipeline, and adds a semantic consistency loss that couples the refined mask to the merged sub-region predictions. The modules are described with enough detail to reimplement, and the ablations in Tables 2 and 3 show monotonic improvements as components are added. That internal consistency is real evidence the design is doing something.\n\nThe soft spots are mostly in the experimental section, and they are serious. First, Table 1 labels [12] as Mask R-CNN, but reference [12] is SegNeXt; [13] is labeled YOLACT but is actually Zeng et al.; [14] is labeled Zeng et al. but is actually Wang et al. That means the reader cannot tell which baseline produced which number. Second, there is no stated training recipe for any baseline—no backbone, optimizer, schedule, or iteration count. Section 4.3 gives BR-Net's settings but not the baselines'. Third, the '200 iterations' with Adam is almost certainly a typo for epochs, which suggests the experimental write-up wasn't carefully checked. Fourth, there are no error bars or repeated runs, and the dataset is private. None of these flaws kill the architectural idea, but they do kill the empirical claim as stated.\n\nWho this is for: people working on C. elegans segmentation or amodal instance segmentation in biological images. A serious referee could help the authors tighten the evaluation and fix the referencing.\n\nRecommendation: Yes, send it to peer review, but with a clear instruction that the baseline protocol and citation errors must be addressed before acceptance. If the authors release code and data and re-run baselines under a common protocol, the paper could become a useful, if modest, contribution.","headline":"A sensible incremental architecture whose empirical superiority claim is not verifiable from the paper as written, due to baseline protocol gaps and citation errors.","tokens_in":13313,"tokens_out":2198,"would_cite":false,"duration_ms":20053,"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":"A bilayer decomposition-recombination network reports higher average precision than prior methods for segmenting overlapping C. elegans on both tested datasets.","keywords":["C. elegans","instance segmentation","amodal instance segmentation","occlusion handling","bilayer segmentation","semantic consistency regularization","attention mechanism","biomedical image analysis"],"falsifier":"Run BR-Net and WormSwin on a public C. elegans instance-segmentation dataset with independently annotated occlusion regions and check whether BR-Net's AP advantage persists; if the gap falls well below the reported differences (0.018 on C.Data-1 and 0.030 on C.Data-2) or reverses, the central claim would be refuted. A simpler adjacent test is to measure inter-annotator agreement on the overlapping/non-overlapping decomposition, since low agreement would show the training targets are not stable enough to support the comparison.","tokens_in":12317,"feed_emoji":"🪱","tokens_out":8639,"duration_ms":75005,"temperature":0.7,"pith_summary":"The paper proposes BR-Net, a deep network for instance segmentation of Caenorhabditis elegans in microscope images where worms overlap and occlude one another. The central claim is that the problem is best handled in two layers: first splitting each worm into overlapping and non-overlapping regions, then recombining those regions under a semantic consistency constraint. On a dataset of 1,092 annotated images split into two subsets, BR-Net reports the highest average precision and mean IoU on both subsets, with AP of 0.739 and 0.712 against 0.721 and 0.682 for the previous best method, WormSwin. If those results hold, the network gives biologists a more accurate tool for tracking individual worms and measuring age-related changes.","feed_headline":"Two-layer network untangles overlapping C. elegans","feed_subtitle":"Splitting worms into occluded and visible regions lifts segmentation accuracy past prior methods on both test sets.","key_machinery":"The load-bearing mechanism is a decompose-recombine loop. The Bilayer Segmentation Module (BSM) uses two parallel mask heads to predict an overlapping mask and a non-overlapping mask for each instance; the Semantic Consistency Recombination Module (SCRM) predicts a refined mask and regularizes it against the XOR-fusion of the two sub-masks, so predicted pixels that appear in both sub-masks are suppressed. The Unified Attention Module (UAM), inserted after the 1x1 convolutions of the feature pyramid, applies channel attention followed by spatial attention to focus the coarse-mask branch on worm instances and away from food residue and excrement in the background. The decomposition gives the network an explicit vocabulary for occlusion, and the XOR consistency constraint keeps the two layers aligned with the final segmentation.","core_discovery":"BR-Net's discovery, as the authors state it, is that occlusion in C. elegans images should be represented explicitly as a bilayer structure rather than left to a single mask head. Its Coarse Mask Segmentation Module first produces a coarse instance mask; its Bilayer Segmentation Module then predicts two sub-masks for each instance, one for the overlapping region and one for the non-overlapping region; and its Semantic Consistency Recombination Module fuses these with region-of-interest features and adds a regularization term that forces the recombined mask to agree with the XOR merge of the two sub-masks. The reported result is that this design exceeds all compared methods on AP, AP50, AP75, and mIoU on both C.Data-1 and C.Data-2, with the largest single jump coming from adding the bilayer decomposition: AP50 rises from 0.671 to 0.821 on C.Data-1.","pith_inferences":["Because the ground-truth sub-masks are derived in-house through positional logic, the magnitude of the reported gains is likely to depend on how the overlapping and non-overlapping regions are defined; a public benchmark with independently annotated occlusion layers would clarify this.","The decompose-recombine design could transfer to other translucent, overlapping biological objects, such as zebrafish larvae or cultured cell colonies, where visible and hidden regions have different visual statistics.","If the bilayer prediction is the main source of the gain, then replacing the XOR merge with a learned fusion that handles overlaps of three or more worms is a natural extension the paper leaves open."],"forward_implications":["If the reported results hold, BR-Net is the best-performing method among the compared approaches on both C.Data-1 and C.Data-2 across all four metrics.","The ablation results attribute the largest single improvement to the bilayer decomposition, with AP50 rising 15 percentage points on C.Data-1 when the Bilayer Segmentation Module is added to the coarse-mask baseline.","The semantic consistency recombination contributes further gains, including 5.3 percentage points of AP50 on the noisier C.Data-2 subset, suggesting the constraint matters most when background clutter is high.","The reported test costs of 2.07 and 2.26 seconds per test are below most amodal baselines, so the accuracy advantage is not bought at a large inference-speed cost."],"supporting_citations":[{"why":"Supplies the Mask R-CNN instance-segmentation baseline whose architecture the Coarse Mask Segmentation Module is built on and whose multitask loss is reused.","marker":"[12]"},{"why":"A worm-specific segmentation method used as a comparison baseline representing prior work on C. elegans instance segmentation.","marker":"[14]"},{"why":"Introduces amodal instance segmentation, the framing the paper adopts, and serves as an early comparison method.","marker":"[17]"},{"why":"Occlusion R-CNN; the paper states its coarse-mask module is inspired by this amodal method, and it is also a baseline.","marker":"[18]"},{"why":"BCNet, the overlapping-bilayer occlusion-aware method that is the closest precursor to the paper's bilayer decomposition and a key baseline.","marker":"[19]"},{"why":"A visible-region-plus-shape-prior amodal segmentation method used as a comparison baseline.","marker":"[20]"},{"why":"The source of the microscope videos from which the two C. elegans datasets are built, so the empirical comparison depends on it.","marker":"[24]"},{"why":"WormSwin, the previous best worm-specific method; the paper's main superiority claim is the AP gap over this method.","marker":"[25]"}],"fun_headline_variants":["Bilayer network separates tangled C. elegans","Two-layer segmentation resolves worm overlap","Split-and-merge network boosts worm segmentation","Explicit bilayer model beats prior worm segmentation","BR-Net untangles overlapping worms with two masks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation stands on the manually drawn instance masks and on the derived overlapping and non-overlapping sub-masks being accurate enough that the reported metric gaps reflect real segmentation ability rather than label noise or arbitrary sub-region definitions.","fun_headline_variants_meta":{"raw":{"variants":["Bilayer network separates tangled C. elegans","Two-layer segmentation resolves worm overlap","Split-and-merge network boosts worm segmentation","Explicit bilayer model beats prior worm segmentation","BR-Net untangles overlapping worms with two masks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000317,"raw_usage":{"total_tokens":1835,"prompt_tokens":1026,"completion_tokens":809,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":742}},"tokens_in":642,"tokens_out":809,"duration_ms":7936,"temperature":1.0,"reasoning_tokens":742,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:58:33.349112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BR-Net and WormSwin on a public C. elegans instance-segmentation dataset with independently annotated occlusion regions and check whether BR-Net's AP advantage persists; if the gap falls well below the reported differences (0.018 on C.Data-1 and 0.030 on C.Data-2) or reverses, the central claim would be refuted. A simpler adjacent test is to measure inter-annotator agreement on the overlapping/non-overlapping decomposition, since low agreement would show the training targets are not stable enough to support the comparison.","supporting_citations":[{"cited_title":"H., Lu, C","cited_arxiv_id":null,"evidence_quote":"Supplies the Mask R-CNN instance-segmentation baseline whose architecture the Coarse Mask Segmentation Module is built on and whose multitask loss is reused."},{"cited_title":"Celeganser: Automated analysis of nematode morphology and age[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops","cited_arxiv_id":null,"evidence_quote":"A worm-specific segmentation method used as a comparison baseline representing prior work on C. elegans instance segmentation."},{"cited_title":"Amodal instance segmentation[C]//Computer Vision –ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11 -14, 2016, Proceedings, Part II 14","cited_arxiv_id":null,"evidence_quote":"Introduces amodal instance segmentation, the framing the paper adopts, and serves as an early comparison method."},{"cited_title":"Learning to see the invisible: End -to-end trainable amodal instance segmentation[C]//2019 IEEE Winter Conference on Applications of Computer Vision (WACV)","cited_arxiv_id":null,"evidence_quote":"Occlusion R-CNN; the paper states its coarse-mask module is inspired by this amodal method, and it is also a baseline."},{"cited_title":"Deep occlusion -aware instance segmentation with overlapping bilayers[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"BCNet, the overlapping-bilayer occlusion-aware method that is the closest precursor to the paper's bilayer decomposition and a key baseline."},{"cited_title":"Amodal segmentation based on visible region segmentation and shape prior[C]//Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"A visible-region-plus-shape-prior amodal segmentation method used as a comparison baseline."},{"cited_title":"WSM -MIL: a weakly supervised segmentation method with multiple instance learning for C elegans image[J]","cited_arxiv_id":null,"evidence_quote":"The source of the microscope videos from which the two C. elegans datasets are built, so the empirical comparison depends on it."},{"cited_title":"WormSwin: Instance segmentation of C","cited_arxiv_id":null,"evidence_quote":"WormSwin, the previous best worm-specific method; the paper's main superiority claim is the AP gap over this method."}],"review_version":1}