{"id":"365dbc15-3132-47b5-89c2-b2efe80e10b6","arxiv_id":"2507.20582","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 2D segmentation network enhanced with a Mesh-Cast sequence module and a shuffle-then-order training schedule reports improved tumor Dice scores on BraTS 2019/2023.","lead":"M-Net treats MRI slices as a temporal sequence and uses a Mesh-Cast module to propagate information through both slice order and imaging channels, keeping a 2D backbone for efficiency. The paper reports accuracy gains on BraTS 2019 and 2023, but the evaluation lacks code, error bars, and an official test protocol.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on a single private split with tiny margins, no error bars, and TPS not applied to baselines; reproducibility is the load-bearing weakness.","rationale":"The reader's weakest assumption identified the evaluation protocol as the key vulnerability: a private split, hidden hyperparameter choices, and no code make the comparison uncheckable. I agree with that diagnosis and add two concrete aggravating details. First, the ablation in Table 3 shows the TPS training strategy, not the Mesh-Cast module, contributes most of the observed gain, so the comparison against baselines trained without TPS does not isolate the claimed architectural contribution. Second, the margins over existing methods are small relative to typical BraTS run-to-run variance, and no repeated runs or error bars are reported; with a single split, differences of 0.1–0.4 Dice cannot be interpreted as reliable superiority. The internal inconsistencies in Table 1 (sequence/slice counts implying non-integer sequence lengths) and Table 4 (invalid timing '97:67') further weaken confidence in the numerical record. These concerns do not invalidate the proposed method, but they do mean the central SOTA claim should remain conditional until code, exact configuration, and repeated-run statistics are provided. Since the reader already returned a CONDITIONAL verdict for essentially this reason, I do not recommend changing the verdict; I would, however, make the release of code and error bars an explicit acceptance condition.","tokens_in":12016,"tokens_out":9642,"duration_ms":113398,"concrete_test":"Release the code, exact split, and training configurations, then rerun Table 4 with five random seeds for M-Net and the strongest baselines (Mamba UNet, Swin UNETR, nnUNet) on the authors' split and on the official BraTS validation split, reporting mean±std for Dice and HD95. Additionally, run the baselines with the TPS phase-one shuffle. If the pairwise margins overlap within one standard deviation, or if the advantage disappears when baselines receive TPS, the Table 4 claim of state-of-the-art performance is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—M-Net achieves state-of-the-art results on every metric in Table 4—depends on one private 8:2 split of the BraTS training sets, with no released code, no fixed random seeds, and no repeated-run statistics. The reported margins over the strongest baselines are very small (e.g., BraTS2019 WT Dice 88.38 vs Mamba UNet 88.21; BraTS2023 ET Dice 93.42 vs Swin UNETR 93.42, a tie). Table 3 shows that much of the gain is attributable to the TPS shuffle phase rather than to the Mesh-Cast architecture: Backbone(Shuffled) already reaches 88.21/90.11/90.86, while full M-Net TPS reaches 88.38/90.52/91.43. Because TPS is not applied to any baseline, the comparison conflates the proposed training strategy with the proposed network. In addition, the numerical record is internally inconsistent: Table 1 implies about 10.33 slices per sequence for BraTS2023 (116250/11250), the sequence length T used in experiments is never stated, and Table 4 contains an invalid inference time for nnUNet ('97:67'). These reporting gaps make the claimed all-metrics superiority unverifiable as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes M-Net, a 2D slice-sequence segmentation framework that treats the MRI slice stack as a temporal sequence. A Mesh-Cast mechanism repeatedly transposes the temporal and channel dimensions so that arbitrary sequential models (LSTM, ConvLSTM, xLSTM, Transformer, Mamba SSM) can process both inter-slice and inter-modality relationships, and a Two-Phase Sequential (TPS) training strategy first trains on globally shuffled sequences then fine-tunes on ordered sequences. The method is evaluated on BraTS2019 and BraTS2023 using an 8:2 split of the official training sets, with Dice and Hausdorff95 metrics, and the authors claim state-of-the-art performance on all regions. The paper also includes ablations across sequential backbones and training phases.","tokens_in":12322,"tokens_out":4224,"duration_ms":44607,"significance":"The conceptual contribution—reinterpreting adjacent MRI slices as a temporal spectrum and using a dimension-swapping mechanism to feed both slice and channel information into any sequence model—is interesting and could offer a computationally cheaper alternative to full 3D convolutions while retaining volumetric context. The TPS training strategy is simple, clearly described, and potentially generalizable to other sequence-based segmentation tasks. The ablation study across five sequential modules is a useful practical comparison. However, the central claim of state-of-the-art performance is not verifiable as presented because the evaluation relies on a private split, all baselines are retrained with unreported hyperparameters, no code or seeds are released, no repeated-run statistics are given, and the reported margins are very small. If the authors make the protocol reproducible and add statistical grounding, the method could be a valuable contribution; in its current form the empirical evidence is insufficient to support the strong SOTA claim.","major_comments":[{"comment":"The state-of-the-art claim rests on a single private 8:2 split of the BraTS training sets, with all comparison methods retrained by the authors under unreported hyperparameters, no fixed random seeds, no repeated runs, and no statistical tests. The margins over the strongest baselines are very small (e.g., BraTS2019 WT Dice 88.38 vs. 88.21 for Mamba UNet; BraTS2023 ET Dice 93.42 vs. 93.42 for Swin UNETR, a tie), so the reported superiority could plausibly be within run-to-run and split-to-split variability. Please release code and exact training configurations for every baseline, report mean±std over multiple seeds, and provide statistical significance tests, or explicitly reposition the results as preliminary.","section":"Section 4.4, Tables 2–4"},{"comment":"The comparison conflates the architecture contribution with the training-strategy contribution because TPS is applied only to M-Net and not to any baseline. Table 3 shows that Backbone(Shuffled) alone already achieves 88.21/90.11/90.86 Dice on BraTS2019, which is comparable to or better than most of the baselines in Table 4; the full M-Net with TPS reaches 88.38/90.52/91.43. To support the claim that the architecture is responsible for the gains, the authors should apply TPS (at least the shuffle phase) to the baseline models, or otherwise separate the architecture and training-strategy effects in the comparison.","section":"Section 4.4, Table 3"},{"comment":"The sequence length T, which is central to the method, is never specified for either dataset, and Table 1 is internally inconsistent. For BraTS2019, 2483+275+702=3460 sequences but the slice counts sum to 51925, which at 155 slices per case corresponds to exactly 335 cases, not 3460 sequences. For BraTS2023 the sequence counts (11250/3750/3763) do not reconcile with the slice counts divided by 155 either. This makes the input construction unverifiable and suggests a mismatch between the reported 'sequences' and actual MRI series. Please state T explicitly for each experiment and correct the data count table.","section":"Section 4.1, Eq. (1), Table 1"},{"comment":"The paper's claim of outperforming 'all key metrics' is contradicted by its own results: on BraTS2023 ET Dice, M-Net (93.42) ties with Swin UNETR (93.42), so the correct summary is 'almost all' metrics. Additionally, the BraTS2019 pattern in which ET Dice (91.43) exceeds WT Dice (88.38) is highly atypical—the enhancing tumor region is normally the most difficult—and the paper offers no explanation for this ordering. Please correct the overclaim and provide an analysis of the unusual region-wise performance pattern, as it may indicate an artifact of the evaluation protocol.","section":"Section 4.4, Table 4, Abstract"}],"minor_comments":[{"comment":"Eq. (23) defines the standard Hausdorff distance, not the 95th percentile Hausdorff distance (HD95) that is used in the experiments; please provide the correct definition of HD95.","section":"Section 4.2, Eq. (23)"},{"comment":"The nnUNet row reports inference time as '97:67', which is not a valid mm:ss format (likely 97 minutes 67 seconds, i.e., 98:07); the UNETR FLOPs value '150.71' is missing the unit (presumably G); and the 'SegResNet' label is inconsistently spaced.","section":"Table 4"},{"comment":"The author email 'sh1yuzh@163.con' contains a typo ('.con' instead of '.com').","section":"Author affiliation block"},{"comment":"The column header 'Dicescore(%)' should be 'Dice score (%)', and the FLOPs column should state the unit for all rows consistently (e.g., 'G').","section":"Table 2"},{"comment":"The notation for sequence length is inconsistent: Eq. (16) uses T, but the text says 'each sequence Xi contain t elements'; please unify the symbols.","section":"Section 3.3"},{"comment":"The caption for Figure 8 reads 'Multi-sequential Module (TPS)' but the text discusses different M-Net configurations; please align the caption with the actual configurations shown.","section":"Figure 8"}],"recommendation":"major_revision","confidential_remarks":"The unusual BraTS2019 result where ET Dice exceeds WT Dice, combined with the private 8:2 split, warrants a check for data leakage or label/preprocessing inconsistencies before we can trust the numerical claims. The self-citation to SLf-UNet (same group) is a minor concern but not a load-bearing issue. If the authors can provide code, seeds, hyperparameters, and multi-run statistics, the paper could become a solid contribution to 2D sequence-based volumetric segmentation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick note on arXiv:2507.20582. The paper has a genuinely usable idea buried under an overclaimed evaluation. The Mesh-Cast mechanism—running a sequence model once over the temporal/slice dimension and once over the channel/modality dimension, with a transpose between—is a simple, clean way to give a 2D segmentation network some 3D context without 3D convolutions. Coupled with the modular \"drop in any sequence model\" design and the TPS two-phase training schedule (shuffle, then ordered fine-tune), that is a real contribution, and the ablations support it. The authors deserve credit for testing several sequence backbones and showing that both temporal and channel modeling add something, and for keeping FLOPs low.\n\nBut the evidence for the headline SOTA claim is thin. Everything rests on a single private 8:2 split of the BraTS training sets, with no released code, no repeated runs, no error bars, and no statistical tests. The margins over the strongest baselines are tiny—0.1–0.5 Dice on BraTS2019, and a flat tie with Swin UNETR on BraTS2023 ET. The ablation in Table 3 shows that much of the apparent gain over Mamba UNet comes from the TPS shuffle phase, but TPS is not applied to any baseline, so you cannot separate the training strategy from the network. There are also reporting gaps: the sequence length T used is never stated (Table 1 implies about 10.3 slices per sequence), the training hyperparameters for baselines are not given, and Table 4 lists an impossible inference time for nnUNet (97:67). None of these are fatal to the method, but they are fatal to the claim as worded.\n\nIf the authors release code, specify the full protocol, report variance over multiple seeds, and re-run baselines with TPS applied, this could become a solid paper. As it stands, it is a promising technical note with an unsupported SOTA statement. I would send it to peer review—the core idea deserves referee time—but I would not accept it in its current form. My verdict would be major revision, and I'd ask for the reproducibility fixes before anything else.","headline":"The Mesh-Cast/TPS ideas are real and worth a look, but the SOTA claim is unsupported as presented because the evaluation rests on one private split with no error bars, tiny margins, and TPS not applied to baselines.","tokens_in":12811,"tokens_out":2544,"would_cite":false,"duration_ms":26719,"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":"M-Net treats MRI slice stacks as temporal sequences and, via a dimension-swapping Mesh-Cast mechanism plus a shuffle-then-order training schedule, reports the best tumor segmentation scores on BraTS2019 and BraTS2023.","keywords":["brain tumor segmentation","MRI slice sequences","temporal-like modeling","Mesh-Cast","sequential image segmentation","Two-Phase Sequential training","Mamba SSM","BraTS"],"falsifier":"Run the released M-Net and all twelve baselines from Table 4 under the paper's exact protocol (8:2 split, 160x160 cropping, Z-score normalization, 300 epochs with early stopping at 30) and check whether M-Net still wins every metric on both datasets; if any baseline reaches or exceeds its Dice or Hausdorff95 values, the universal-superiority claim fails.","tokens_in":11831,"feed_emoji":"🧠","tokens_out":5080,"duration_ms":50593,"temperature":0.7,"pith_summary":"This paper tries to establish that the spatial continuity of adjacent MRI slices can be exploited as if it were time, and that a 2D segmentation network equipped with a dimension-swapping wrapper, Mesh-Cast, can capture that 'temporal-like' context without the computational cost of full 3D convolutions. M-Net interleaves a sequence model across the slice dimension and across the feature-channel dimension, then trains in two phases: first on globally shuffled slice sequences, then on the true slice order. On BraTS2019 and BraTS2023, the authors report that M-Net with Mamba SSM outperforms all twelve compared methods on every Dice and Hausdorff95 metric for whole tumor, tumor core, and enhancing tumor. If the comparison is fair and reproducible, the practical consequence is that volumetric context in MRI segmentation does not require a 3D backbone.","feed_headline":"MRI slices as a sequence: M-Net tops tumor segmentation","feed_subtitle":"A mesh-based dimension swap lets 2D networks use 3D context, beating every baseline on BraTS 2019 and 2023.","key_machinery":"Mesh-Cast is a transpose-and-sweep wrapper around any Sequential Module. Input features have shape T frames by C channels by D flattened pixels; the module first runs the sequence model with T as the sequence length and C treated like a batch dimension, then swaps the T and C roles with a transpose, runs the same sequence model along the channel axis, and transposes back. This repeated mesh of temporal and channel sweeps lets correlations across slices and across MRI modalities both be captured. The Two-Phase Sequential training strategy first trains on global frame-level shuffles and then fine-tunes on ordered sequences, which the paper argues lets the model learn common lesion patterns before sequence-specific continuity.","core_discovery":"The central claim is that treating a stack of T MRI slices as a sequence, and pushing a sequential model first along the slice axis and then along the channel axis, is sufficient to reach the top of brain tumor segmentation benchmarks. On BraTS2019, M-Net reports Dice scores of 88.38 for whole tumor, 90.52 for tumor core, and 91.43 for enhancing tumor, with corresponding Hausdorff95 distances of 1.2869, 0.8154, and 0.6571; on BraTS2023, it reports 91.33, 93.55, and 93.42, leading every baseline on both datasets. The paper attributes the gain to the Mesh-Cast dual-axis sweep rather than to any particular sequence model, since the same wrapper improves LSTM, ConvLSTM, xLSTM, Transformer, and Mamba variants over their slice-only backbones by up to 1.2 Dice points.","pith_inferences":["Mesh-Cast is potentially a general plug-in for any consistent-pattern sequence of images, such as CT slice stacks, video frames, or multi-view scans; the paper states this generality but does not test it.","The channel-axis sweep may be doing more of the work than the paper's ablation fully isolates, because co-registered MRI modalities make channel-order correlations a proxy for inter-modal alignment; a test on single-modality sequences would separate the two.","The TPS shuffle phase could act as a data augmentation that increases effective dataset diversity, which would predict larger gains on small datasets and diminishing returns as data grows; this is testable by scaling the training set.","A direct ablation that replaces Mesh-Cast with simple concatenation of adjacent slices would clarify whether the dual-axis sweep itself, rather than the added sequence model, drives the reported gains."],"forward_implications":["Plugging Mesh-Cast into an existing 2D encoder-decoder should produce immediate gains on slice-stack medical data, because every tested sequence model improved over the slice-only backbone.","Sequence models become interchangeable components: the best variant uses Mamba SSM, but LSTM, ConvLSTM, xLSTM, and Transformer variants all improve, so future backbones can swap in newer sequence models without redesign.","Volumetric segmentation becomes feasible on limited hardware: M-Net's 91.29 GFLOPs and roughly 15-minute inference time compare with nnU-Net's 82 GFLOPs and roughly 97-minute inference time under the paper's setup.","The TPS schedule is a separable contribution: phase-one shuffling alone already beats the ordered backbone on most metrics, and the full shuffle-then-order schedule is better than either phase alone."],"supporting_citations":[{"why":"Supplies the U-Net encoder-decoder backbone that M-Net extends with Mesh-Cast sequential modules.","marker":"[29]"},{"why":"Defines the Mamba SSM sequential model used in the best-performing M-Net configuration.","marker":"[13]"},{"why":"Defines xLSTM, one of the interchangeable sequential modules evaluated in the ablation study.","marker":"[5]"},{"why":"Defines LSTM, the basic sequence model that the Mesh-Cast wrapper generalizes.","marker":"[12]"},{"why":"Provide the BraTS 2019 and BraTS 2023 datasets, modalities, labels, and benchmark context used for training and evaluation.","marker":"[3, 4, 26]"},{"why":"Provides nnU-Net, the strongest 3D baseline and the main inference-time comparison.","marker":"[18]"},{"why":"Provides Swin UNETR, a state-of-the-art transformer baseline that M-Net must outperform.","marker":"[16]"},{"why":"Provides Mamba UNet, the sequence-model baseline closest in spirit to M-Net.","marker":"[34]"}],"fun_headline_variants":["M-Net's Mesh-Cast makes MRI slices a winning sequence","Sequence-aware Mesh-Cast tops brain tumor segmentation","MRI slices as sequence: M-Net beats all on BraTS","Mesh-Cast dual sweep gives 2D nets top tumor scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation protocol is fair and reproducible: all methods were retrained on the same 8:2 split, the same cropping and normalization, the same loss, and the same early-stopping rule, with no hidden hyperparameter tuning that favours M-Net.","fun_headline_variants_meta":{"raw":{"variants":["M-Net's Mesh-Cast makes MRI slices a winning sequence","Sequence-aware Mesh-Cast tops brain tumor segmentation","MRI slices as sequence: M-Net beats all on BraTS","Mesh-Cast dual sweep gives 2D nets top tumor scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":3073,"prompt_tokens":989,"completion_tokens":2084,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":2013}},"tokens_in":605,"tokens_out":2084,"duration_ms":16484,"temperature":1.0,"reasoning_tokens":2013,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:39:52.039289+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released M-Net and all twelve baselines from Table 4 under the paper's exact protocol (8:2 split, 160x160 cropping, Z-score normalization, 300 epochs with early stopping at 30) and check whether M-Net still wins every metric on both datasets; if any baseline reaches or exceeds its Dice or Hausdorff95 values, the universal-superiority claim fails.","supporting_citations":[{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net encoder-decoder backbone that M-Net extends with Mesh-Cast sequential modules."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines xLSTM, one of the interchangeable sequential modules evaluated in the ablation study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines LSTM, the basic sequence model that the Mesh-Cast wrapper generalizes."},{"cited_title":"Jaeger, Simon A","cited_arxiv_id":null,"evidence_quote":"Provides nnU-Net, the strongest 3D baseline and the main inference-time comparison."},{"cited_title":"Hatamizadeh, V","cited_arxiv_id":null,"evidence_quote":"Provides Swin UNETR, a state-of-the-art transformer baseline that M-Net must outperform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Mamba UNet, the sequence-model baseline closest in spirit to M-Net."}],"review_version":2}