{"id":"369e4f9f-d4be-428b-9634-f6deb00e0056","arxiv_id":"1908.08071","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A boundary-aware CNN with a dedicated edge stream and edge losses improves tumor-core segmentation on 2D slices of BraTS 2018 compared to U-Net and V-Net, though the evaluation has internal inconsistencies.","lead":"This paper adds an edge-detection branch to a standard image segmentation network and trains it with extra boundary-focused losses, then tests it on brain tumor MRI slices. The authors report higher Dice scores than two older baseline networks, but the reported numbers contain a mathematical inconsistency and the test setup is underspecified.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 reports Dice and Jaccard values that are mathematically incompatible, so the central quantitative claim is unsupported as written.","rationale":"The reader's weakest_assumption focused on the unspecified 2D slice-based evaluation protocol and its potential lack of generalizability. I agree that this is a serious issue. However, the more load-bearing problem is the internal inconsistency in the reported metrics: the Dice and Jaccard values in Table 1 cannot both be correct under the standard definitions used in the paper. Since the central claim is a quantitative superiority claim, this directly undermines the evidence. The reader's rationale did mention 'internal inconsistency in the reported metrics' but did not make it the primary weak assumption. I recommend UNVERDICTED rather than CONDITIONAL because the current submission does not provide internally consistent quantitative evidence, and the missing split prevents a meaningful verification. The proposed test is straightforward: recompute the metrics from the actual predictions; if the inconsistency is confirmed, the paper must be revised with corrected metrics and a properly described patient-disjoint evaluation before the central claim can be assessed.","tokens_in":4536,"tokens_out":4779,"duration_ms":49949,"concrete_test":"Reproduce Table 1 from the saved prediction and ground-truth masks using Eq. (4) for Dice and Jaccard = |A∩B|/|A∪B|. If any row violates the necessary inequality Jaccard ≤ Dice, the published metrics are invalid. Then re-compute all rows on a patient-disjoint train/test split to determine whether the Dice margin of 0.822 vs 0.769 (V-Net) and 0.731 (U-Net) survives.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim that the boundary aware network outperforms U-Net and V-Net rests entirely on Table 1. Under the standard definitions invoked in §3.2 and Eq. (4), for any two binary masks A and B, Dice = 2|A∩B|/(|A|+|B|) and Jaccard = |A∩B|/|A∪B|, with |A∪B| ≤ |A|+|B|, so Dice ≥ Jaccard always. Table 1 reports Dice 0.731, 0.769, 0.768, 0.822 and Jaccard 0.805, 0.837, 0.832, 0.861; every Jaccard exceeds its corresponding Dice, which is impossible. At least one metric is computed with a nonstandard definition or the table entries are misreported. The ablation comparison (Ours no edge loss, Dice 0.768, vs Ours, Dice 0.822) is affected by the same inconsistency, so the magnitude of the claimed improvement cannot be trusted. Additionally, §3.1 does not describe a train/test split, patient-level separation, or slice count for the 2D dataset derived from BraTS, allowing possible leakage; however, the metric inconsistency alone is sufficient to invalidate the reported quantitative comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end trainable boundary-aware convolutional network for medical image segmentation. The architecture augments an encoder-decoder segmentation stream with a shape stream that receives features from each encoder resolution, applies attention layers, and produces boundary predictions; the training loss combines Dice losses on the segmentation and boundary outputs with a weighted binary cross-entropy edge loss. The method is evaluated on a 2D axial-slice dataset built from BraTS 2018 volumes for tumor core segmentation and compared with U-Net and V-Net. The authors report Dice, Jaccard, and Hausdorff distances and conclude that the boundary-aware network outperforms the baselines, with an ablation showing that the edge loss is responsible for most of the gain.","tokens_in":4829,"tokens_out":6096,"duration_ms":55873,"significance":"The idea of explicitly supervising an edge stream alongside a segmentation stream is sensible and has been explored in related work; the paper's contribution is the specific architecture (attention layers, dilated spatial pyramid pooling, combined losses) and its application to BraTS. Strengths include an end-to-end trainable design and a direct ablation of the proposed edge loss, which is a legitimate comparison. However, the quantitative evidence is currently unreliable: the reported Dice/Jaccard pairs are mathematically inconsistent, the evaluation protocol lacks a described train/test split and patient-level separation, and no significance tests are reported. If corrected, the ablation could still support the mechanism, but the paper as written does not establish the claimed margin over U-Net and V-Net.","major_comments":[{"comment":"The reported Dice and Jaccard values are mutually incompatible under the standard definitions used in the paper. For any two sets, Dice >= Jaccard because |A∪B| <= |A|+|B|; however, Table 1 reports Dice 0.731, 0.769, 0.768, 0.822 with Jaccard 0.805, 0.837, 0.832, 0.861, so every Jaccard value exceeds its corresponding Dice. Since this relation is impossible, at least one metric was computed with a nonstandard definition or the entries are misreported. The central comparison and the edge-loss ablation both depend on these numbers and must be corrected or recomputed.","section":"Table 1, §3.2 Eq. (4)"},{"comment":"The evaluation protocol is underspecified and likely biased. The authors state that \"several axial slices centered around the tumor region\" were extracted from each 3D volume and combined into a 2D dataset, but they do not report the number of slices, how train/validation/test sets were split, or whether splits were performed at the patient/volume level. Without patient-level separation, slices from the same volume can appear in both training and testing, allowing leakage; with tumor-centered slice selection, the evaluation is restricted to slices with large tumor burden and does not measure performance on full 3D volumes. The reported standard deviations are large and overlap across methods, and no significance tests are given, so the claimed improvement over V-Net and U-Net is not statistically supported.","section":"§3.1, Table 1"},{"comment":"Equation (5) does not define the edge loss as introduced. In Eq. (3), LEdge is a function of (spred, strue), but Eq. (5) uses ypred,j in the probability terms, which refers to the main-stream segmentation output, and the definition of the sets y+ and y- is incomplete (they are said to be edge and non-edge pixel sets, but their dependence on strue is not stated). As written, the equation cannot be implemented unambiguously, and it is unclear whether the boundary stream is supervised as claimed. Please rewrite the loss with explicit arguments and define y+ and y- in terms of strue.","section":"§2.3, Eq. (5)"},{"comment":"The comparison with U-Net and V-Net is not fully specified. The paper does not state the loss function, optimizer settings, number of epochs, input/output configuration, or data augmentation used for the baselines, nor whether they were trained on the same 2D slices with the same hyperparameters. Without these details, the reported differences could reflect implementation choices rather than architectural benefit. Please provide the training protocol for all models.","section":"§3.2, Table 1"}],"minor_comments":[{"comment":"There is a typo in \"nectrotic core\" which should be \"necrotic core.\"","section":"§3.1"},{"comment":"The phrase \"probability of the predicated class\" contains a typo and should read \"predicted class.\"","section":"§2.3, Eq. (5)"},{"comment":"The initial learning rate is written as α0 = 1.0−3, which should be formatted as 1e-3 or 10^-3 for clarity.","section":"§3.2, Eq. (6)"},{"comment":"The caption says the ablation is performed \"with and without this layer,\" but the text describes ablating the edge loss while keeping the architecture; this discrepancy should be clarified.","section":"Table 1 caption"},{"comment":"The statement that strue \"can be obtained by computing the spatial gradient of ytrue\" is ambiguous; please specify the exact edge-extraction method (e.g., morphological gradient, Sobel, Canny) and any thresholding or post-processing.","section":"§2.3"},{"comment":"The Hausdorff distance is not defined; please state whether it is the 95th percentile Hausdorff distance and in which units it is reported.","section":"§3.2, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The Table 1 inconsistency should be treated as a data-integrity check rather than assumed misconduct; the authors must provide corrected metrics and a detailed evaluation protocol. The paper is better suited to a workshop or short conference format unless the experiments are substantially strengthened with proper patient-level splits, full 3D evaluation, and statistical testing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'd keep this paper on your radar, but not for the reported numbers. The architecture is a sensible transfer of edge-aware shape streams (Gated-SCNN style) to medical image segmentation, and the ablation design—same architecture with and without the edge loss—is exactly the right experiment to run. The attention layers in the shape stream are a reasonable addition, and the qualitative figures suggest the boundary emphasis is working visually.\n\nThe problem is Table 1. The paper defines Dice and Jaccard in the standard way (Eq. 4 and §3.2), and under those definitions Dice ≥ Jaccard always. But every row reports a Jaccard larger than its Dice: 0.805 vs 0.731, 0.837 vs 0.769, and so on. That is mathematically impossible. At least one metric is computed with a nonstandard definition or the table entries are misreported. Since the central claim—that the edge loss improves Dice from 0.768 to 0.822—rests entirely on that table, the quantitative comparison is unsupported as written.\n\nThe evaluation protocol is also under-specified. The authors extract 2D axial slices \"centered around the tumor region\" from 3D BraTS volumes, but give no slice count, no train/test split, and no patient-level separation. That risks leakage or bias, especially when comparing against baselines that may not be tuned on the same slice distribution. They also report no significance tests and compare only against U-Net and V-Net, without any modern strong baseline. The self-citations are minor and not circular; that's not an issue.\n\nHaving said that, the core idea is not invalid. Edge supervision as a regularizer for medical segmentation is a legitimately useful training mechanism, and the ablation, once the metrics are fixed, could still show a real effect. The paper just needs a corrected results table, an honest description of the data split, and a stronger baseline comparison before the effect size can be trusted.\n\nWho is this for? Readers working on boundary-aware segmentation in medical imaging, and possibly a MICCAI-style workshop audience. It deserves a serious referee—not a desk reject—but the referee should treat the quantitative claims with skepticism until the errors are fixed. I'd recommend accepting it into peer review with major revision clearly required.","headline":"A plausible boundary-aware segmentation mechanism undermined by an internally inconsistent results table and an under-specified evaluation protocol.","tokens_in":5326,"tokens_out":1435,"would_cite":false,"duration_ms":16691,"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 boundary-aware network with an edge-loss branch segments brain tumors more accurately than U-Net or V-Net, lifting Dice from 0.731/0.769 to 0.822.","keywords":["medical image segmentation","semantic segmentation","boundary aware networks","edge-aware loss","brain tumor segmentation","encoder-decoder CNN","BraTS 2018","Dice loss"],"falsifier":"Retrain the full model and the no-edge-loss ablation on complete 3D BraTS volumes using the standard training/validation split and compare whole-tumor-core Dice; if the gap between the two models falls to near zero, the reported edge-loss benefit is an artifact of the 2D slice selection rather than a genuine segmentation improvement.","tokens_in":4379,"feed_emoji":"🧠","tokens_out":7961,"duration_ms":65696,"temperature":0.7,"pith_summary":"This paper argues that teaching a segmentation network where organ boundaries are—through a dedicated edge branch and edge-aware loss terms—yields more accurate medical image segmentation than standard U-Net and V-Net encoders. On brain tumor segmentation from the BraTS 2018 dataset, the full boundary-aware model reaches a Dice score of 0.822, versus 0.769 for V-Net and 0.731 for U-Net; Dice is a standard overlap measure between predicted and labeled regions. The paper also shows the gain comes from the edge loss, not just the extra branch: removing the edge loss drops the score to 0.768, essentially matching V-Net. If true, this offers an end-to-end trainable way to improve segmentation accuracy by making networks attend to boundary structure rather than texture alone.","feed_headline":"Edge-aware model tops U-Net and V-Net on tumor Dice","feed_subtitle":"A dedicated edge stream plus edge loss is the source of the gain; removing it drops Dice from 0.822 to 0.768.","key_machinery":"The mechanism is a two-stream encoder-decoder: a main residual-block segmentation stream plus a shape stream that receives encoder feature maps at each resolution, applies a $1\\times 1$ convolution, and passes them through attention layers that compute $\\alpha_l = \\sigma(C_{1\\times1}(s_l \\| m_l))$ and gate the shape features by element-wise multiplication. The shape stream's boundary prediction is supervised by a generalized Dice loss plus a weighted binary cross-entropy edge loss, and its last attention output is concatenated with the encoder output and fed into dilated spatial pyramid pooling. The edge loss is the load-bearing component: with it, Dice reaches 0.822; without it, the boundary stream adds nothing and the model behaves like V-Net.","core_discovery":"The central discovery is that supervising an auxiliary edge-stream branch with a weighted binary cross-entropy edge loss, combined with Dice losses on both the main segmentation output and the edge output, regularizes the shared encoder so that it produces finer boundaries and fewer false positives. The authors claim that their boundary-aware network outperforms U-Net and V-Net across Dice score, Jaccard index, and Hausdorff distance on the tumor-core segmentation task, and that the ablation without the edge loss performs nearly identically to V-Net (0.768 versus 0.769 Dice), pinning the improvement on the edge-aware supervision rather than on the extra parameters of the shape stream.","pith_inferences":["A natural extension the paper leaves untested is applying the same edge-supervised stream to full 3D volumes; because the paper deliberately reduces BraTS to 2D slices centered on the tumor, the reported gains might shrink or grow when evaluated on whole-volume segmentation.","The texture-versus-shape motivation suggests the edge-aware loss might also improve robustness to image appearance shifts, a property that could be tested by evaluating on a different MRI protocol or on natural images.","The absence of a described train/test split and slice count means the quantitative advantage should be re-measured on a standard 3D BraTS validation split before relying on the exact numbers; this is an editorial caution, not a claim in the paper.","One could combine the edge-loss idea with other shape-bias mechanisms, such as explicit boundary alignment, to see whether the gains accumulate."],"forward_implications":["The edge-aware loss is the active ingredient: the same architecture without it gives a Dice of 0.768, essentially matching V-Net's 0.769.","Boundary-aware training cuts false positives and sharpens edges, improving all three reported metrics over both baselines.","The gain is available under a simplified setting (single T1c modality, single tumor core region, 2D slices), so it does not depend on using all four BraTS modalities.","Since the edge stream is supervised mainly during training, adding it does not complicate inference compared with the baseline encoder-decoder."],"supporting_citations":[{"why":"Provides the U-Net baseline architecture and the main comparison that the boundary-aware model outperforms.","marker":"[9]"},{"why":"Provides the V-Net baseline and the generalized Dice loss used in both the main and edge streams.","marker":"[7]"},{"why":"Supplies the BraTS 2018 dataset with expert tumor annotations used for training and evaluation.","marker":"[2]"},{"why":"Motivates the boundary-aware design by showing CNNs are biased toward texture rather than shape.","marker":"[3]"},{"why":"Supplies the learning-rate schedule and residual-block design conventions the training follows.","marker":"[8]"}],"fun_headline_variants":["Edge stream boosts brain tumor segmentation Dice","Boundary-aware net beats U-Net and V-Net on tumors","Edge loss sharpens tumor boundaries in MRI","Auxiliary edge branch lifts tumor Dice to 0.822"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation rests on 2D axial slices extracted from 3D BraTS volumes without a described train/test split or slice count, so the reported Dice gains assume those slices represent realistic full-volume tumor segmentation.","fun_headline_variants_meta":{"raw":{"variants":["Edge stream boosts brain tumor segmentation Dice","Boundary-aware net beats U-Net and V-Net on tumors","Edge loss sharpens tumor boundaries in MRI","Auxiliary edge branch lifts tumor Dice to 0.822"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1188,"prompt_tokens":801,"completion_tokens":387,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":417,"completion_tokens_details":{"reasoning_tokens":322}},"tokens_in":417,"tokens_out":387,"duration_ms":3581,"temperature":1.0,"reasoning_tokens":322,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:49:54.718953+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the full model and the no-edge-loss ablation on complete 3D BraTS volumes using the standard training/validation split and compare whole-tumor-core Dice; if the gap between the two models falls to near zero, the reported edge-loss benefit is an artifact of the 2D slice selection rather than a genuine segmentation improvement.","supporting_citations":[{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Provides the U-Net baseline architecture and the main comparison that the boundary-aware model outperforms."},{"cited_title":"In: Fourth International Conference on 3D Vision (3DV) (2016)","cited_arxiv_id":null,"evidence_quote":"Provides the V-Net baseline and the generalized Dice loss used in both the main and edge streams."},{"cited_title":"Scientific Data 4 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the BraTS 2018 dataset with expert tumor annotations used for training and evaluation."},{"cited_title":"In: International Conference on Learning Representations (ICLR) (2019)","cited_arxiv_id":null,"evidence_quote":"Motivates the boundary-aware design by showing CNNs are biased toward texture rather than shape."},{"cited_title":"In: BrainLes, Medical Image Computing and Computer Assisted Intervention (MICCAI)","cited_arxiv_id":null,"evidence_quote":"Supplies the learning-rate schedule and residual-block design conventions the training follows."}],"review_version":1}