{"id":"19cc452c-9383-4b48-9300-46df9f1e5f7a","arxiv_id":"2504.21281","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Mamba-based architecture with modality-specific encoders and adaptive multi-level feature fusion reports strong tumor segmentation results on BraTS2023 and Hecktor2022 benchmarks.","lead":"This paper introduces a Mamba-based deep learning model that segments tumors in 3D medical images by processing each imaging modality separately and then adaptively fusing their features. The model reports high Dice scores on brain tumor MRI and head and neck PET/CT datasets, and in the tables it outperforms several established CNN, Transformer, and Mamba approaches.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed BraTS2023 state of the art in Section 3.4 rests on Table 2 baselines imported from SegMamba [19] under an uncontrolled split and preprocessing; the 0.83 Dice / 0.94 mm margins are unestablished until a common protocol is re-run.","rationale":"The paper proposes a coherent Mamba-based multi-modal segmentation architecture, and the stress-test should be read in that light. The method is clearly specified (specific-modality Mamba encoders, bi-level synergistic integration block, decoder with skip connections and an extra Mamba block at the bottleneck), and the internal ablation in Table 1 is internally consistent: Ours > MambaEncoder > SimpleFusion > SingleModality on every metric, with the stated 2.02 Dice / 1.16 mm HD gains matching the table. That is real supporting evidence for the architecture's internal design choices. The load-bearing weakness is exclusively at the level of the external claim. Section 3.4 asserts the model 'outperforms all segmentation approaches,' but the evidence is Table 2, where the baseline rows are imported from [19] (as Section 3.3 explicitly admits) without re-running them under the authors' 70/10/20 split, preprocessing, or metric definition. Because the margins over the strongest baseline (SegMamba) are only 0.83 mean Dice and 0.94 mm HD, and no variance or significance information is reported, the superiority claim is indistinguishable from protocol noise as presented. I also flag Eq. 5: it defines the maximum (100th percentile) Hausdorff distance, whereas [19] and most BraTS papers report HD95; if the variants are mixed, the HD column is biased. This reinforces, rather than replaces, the primary split/preprocessing concern. The reader's weakest_assumption identifies exactly this uncontrolled baseline comparison, and I agree with that assessment; my read does not move the verdict. CONDITIONAL remains appropriate: the architecture and internal ablations warrant publication if the authors supply a controlled re-comparison (or evaluate on the official BraTS2023 validation set with the same metric) and report error bars. I therefore recommend no change to the reader's verdict.","tokens_in":8141,"tokens_out":9516,"duration_ms":88748,"concrete_test":"Re-run the proposed model and the public SegMamba implementation under one common protocol: identical BraTS2023 70/10/20 split, identical preprocessing and patch sampling, at least three seeds, and one stated Hausdorff variant (max-HD per Eq. 5 or HD95). If the mean Dice and HD margins over SegMamba shrink below the reported 0.83 / 0.94 or reverse sign, the Section 3.4 'outperforms all' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, 'our approach outperforms all segmentation approaches' (Section 3.4), is supported only by Table 2, where six of seven baseline rows are quoted, not re-run. Section 3.3 states: 'For the BraTS2023 dataset, the evaluation metrics for comparison methods are sourced from [19].' The authors test on their own random 70/10/20 split (Section 3.3), whereas SegMamba's published numbers come from SegMamba's own protocol, with different test cases, preprocessing, and patch/crop handling. The reported edge over SegMamba — 0.83 mean Dice (92.15 vs 91.32) and 0.94 mm HD (2.62 vs 3.56) — is small enough to be explained by split or preprocessing differences, and no error bars, seeds, or significance test are given. A second comparability hazard: Eq. 5 defines Hausdorff distance as the true maximum over boundary points, while [19] and most BraTS papers report the 95th percentile HD; if the variants differ, the HD column is not comparable. The internal ablation (Table 1) is self-consistent and each added component helps, but it does not validate the external comparison. The Hecktor2022 experiment (Table 3) is only versus a PET-only variant of the same network, so it supports multi-modal fusion but not any state-of-the-art claim. The concern is testable because SegMamba's code is public.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-modal 3D tumor segmentation architecture built on Mamba-based encoders. Each input modality is processed by a dedicated Mamba encoder to extract long-range features, and a 'bi-level synergistic integration block' fuses modalities using modality-level and channel-level attention. A decoder with skip connections produces the segmentation map. The method is evaluated on BraTS2023 (MRI, four modalities) and Hecktor2022 (PET/CT). The authors report ablation results showing incremental gains from each component, and a comparison table on BraTS2023 in which their method outperforms seven published CNN, Transformer, and Mamba baselines. On Hecktor2022, they compare their PET/CT fusion against a PET-only variant of the same network.","tokens_in":8448,"tokens_out":3212,"duration_ms":32805,"significance":"If the reported performance is reproducible, the architecture is a sensible and potentially useful contribution to multi-modal 3D segmentation: the modality-specific Mamba encoders and the adaptive fusion block are well motivated, and the internal ablation in Table 1 is self-consistent and shows a clear benefit from each added component. The paper also demonstrates the value of PET/CT fusion over PET-only input on Hecktor2022. However, the central claim of state-of-the-art performance on BraTS2023 is currently not supported with sufficient rigor, because the baseline numbers are imported from a single prior paper under a different experimental protocol, and the reported margins are small. The manuscript's significance therefore depends on whether the external comparison can be placed on a common footing.","major_comments":[{"comment":"The claim 'our approach outperforms all segmentation approaches' in Section 3.4 rests on Table 2, but six of the seven baseline rows are quoted from SegMamba [19] rather than re-run under the authors' own protocol. The authors state in Section 3.3: 'For the BraTS2023 dataset, the evaluation metrics for comparison methods are sourced from [19].' The authors train on their own random 70/10/20 split of BraTS2023, whereas the SegMamba numbers were obtained under a different split and preprocessing pipeline. The reported margins are small (0.83% mean Dice and 0.94 mm HD over SegMamba) and could plausibly be explained by split, preprocessing, or patch/crop differences. No error bars or significance tests are given. To support the state-of-the-art claim, the authors should re-run all baselines under the same split, preprocessing, and evaluation metric, or use the official BraTS2023 validation/test protocol, and report variance across multiple seeds.","section":"Section 3.3, Table 2, Section 3.4"},{"comment":"The Hausdorff distance in Eq. (5) is defined as the maximum Euclidean distance between boundary points, but most BraTS papers, including SegMamba [19], report the 95th percentile Hausdorff distance (HD95). If the baseline numbers in Table 2 are HD95 and the authors' numbers are maximum HD, then the HD column is comparing two different quantities, and the reported improvements are not meaningful. The authors must state explicitly which variant of HD is used for their own results and for each baseline, and ensure that all numbers in Table 2 are computed with the same metric (ideally by re-running the baselines).","section":"Section 2 (Eq. 5) and Table 2"},{"comment":"The Hecktor2022 experiment in Table 3 compares only the authors' PET/CT model against a PET-only variant of the same network. This demonstrates that adding CT information improves segmentation within their architecture, but it does not support the abstract's claim of 'competitive performance compared to the state-of-the-art CNN, Transformer, and Mamba-based approaches' on this dataset. The authors should either add comparisons with published Hecktor2022 results under a common evaluation protocol, or explicitly restrict the claim on Hecktor2022 to demonstrating the benefit of multi-modal fusion rather than state-of-the-art performance.","section":"Section 3.4, Table 3"}],"minor_comments":[{"comment":"The sentence 'The structure of the paper is as follows: Section?? reviews the state-of-the-art segmentation methods' contains an unresolved placeholder 'Section??' and must be completed.","section":"Section 1, final paragraph"},{"comment":"There are grammatical and typographical errors, including 'an bi-level' in the abstract and Section 2.2, and 'quantifies the the maximum boundary deviation' in Section 3.2. These should be corrected.","section":"Throughout"},{"comment":"The description of the concatenation in the bi-level integration block is ambiguous: it is unclear whether the M modality feature maps are concatenated into a new modality dimension or along the channel dimension. The equations for the attention weights would benefit from explicit tensor shapes.","section":"Section 2.2"},{"comment":"Implementation details are incomplete: the paper does not state the input patch size, cropping strategy, data augmentation, or the exact composition of the loss (cross-entropy only). Figure 1 is too small to read the individual block labels, which are crucial for understanding the architecture.","section":"Section 3.3 and Figure 1"},{"comment":"The caption of Table 2 should explicitly state that the baseline numbers are quoted from [19] and were not re-run by the authors under their own protocol, to avoid misleading readers.","section":"Table 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable engineering contribution with a clean ablation, but the flagship claim of state-of-the-art performance on BraTS2023 is not supported by the current comparison methodology. The authors need to either re-run the baselines under a common protocol (feasible, since SegMamba's code is public) or substantially temper the claims. I would also note that the paper's novelty relative to SegMamba is incremental and the writing needs polishing; however, the core architectural idea is sound and the ablation is encouraging."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible architecture paper, but the headline claim of state-of-the-art on BraTS2023 is not backed up by the experiments as reported. The method is a sensible combination of per-modality Mamba encoders, a bi-level attention fusion block, and a decoder. That combination appears new, and the ablation in Table 1 is internally consistent: each component adds a clear increment. The writing is understandable once you ignore the formatting glitches like 'Section??' and 'F. Author et al.'\n\nThe main problem is the external comparison. Table 2 lists seven baselines, but the numbers for those baselines are taken from SegMamba [19], not re-run under the authors' own 70/10/20 split and preprocessing. The authors say so in Section 3.3. That makes the claimed edge over SegMamba (0.83 Dice, 0.94 mm HD) fragile; the margin is exactly the size that split or preprocessing differences can produce. There are no error bars, seeds, or significance tests.\n\nThere is an additional comparability hazard that the stress-test note caught, and I think it is real: Eq. 5 defines Hausdorff distance as the true maximum, while SegMamba and most BraTS papers report the 95th percentile. If those are different quantities, the HD column in Table 2 is not comparing like with like. The authors need to clarify this.\n\nThe Hecktor2022 experiment (Table 3) only compares the full model against its PET-only variant. That supports the value of fusion but not any state-of-the-art claim. Fine, but it is not evidence for 'outperforms all approaches.'\n\nOn substance: the architecture is incremental over SegMamba and attention fusion, but the modality-specific encoder idea is worth exploring, and the ablation suggests it helps. No theoretical contribution, no released code, no statistical validation. The Mamba equations are standard from prior work.\n\nWho is this for? People working on Mamba-based medical segmentation will find it a useful related-work reference and a reasonable baseline to compare against. It is not a field-changer.\n\nMy recommendation: send it to peer review, but with a clear request for a controlled comparison — either re-run baselines under the same protocol or use the official BraTS validation/leaderboard, and report the same HD variant for all methods. Without that, the central claim does not hold. If the authors provide that, the paper could be acceptable as a modest contribution.","headline":"A plausible Mamba-fusion architecture whose BraTS2023 SOTA claim is not established because the baseline numbers are borrowed and the Hausdorff metric may be inconsistent.","tokens_in":8961,"tokens_out":3023,"would_cite":true,"duration_ms":29595,"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 feeding each 3D modality through its own Mamba encoder and fusing the results with two-level attention beats seven CNN, Transformer, and Mamba baselines on BraTS2023 and improves PET/CT fusion over PET-only on…","keywords":["3D medical image segmentation","tumor segmentation","Mamba","multi-modal feature fusion","bi-level attention","BraTS2023","PET/CT","state space model"],"falsifier":"Re-run the seven compared methods under the exact same BraTS2023 split and preprocessing used in this paper, then test the mean Dice difference between this method and SegMamba, 92.15 versus 91.32, with a paired test across test cases; if the difference is not significant or reverses, the central outperformance claim fails.","tokens_in":7945,"feed_emoji":"🧠","tokens_out":4972,"duration_ms":47700,"temperature":0.7,"pith_summary":"The paper argues that 3D multimodal tumor segmentation improves when each imaging modality is processed by its own Mamba encoder before fusion, rather than stacking modalities as input channels. It introduces a bi-level synergistic integration block that learns modality-level and channel-level attention weights to combine features adaptively. On BraTS2023 MRI, the method reports a mean Dice score of 92.15 and a mean Hausdorff distance of 2.62 mm, beating the seven CNN, Transformer, and Mamba baselines listed in the comparison. On Hecktor2022 PET/CT, fusing both modalities raises the Dice score from 33.07 with PET-only input to 37.12. If the comparisons hold, the practical lesson is that explicit per-modality encoding plus learned attention fusion is a better use of Mamba than a single shared encoder.","feed_headline":"Mamba plus adaptive fusion beats baselines in 3D tumor segmentation","feed_subtitle":"Two-level attention fusion lifts BraTS2023 Dice to 92.15 and PET/CT Dice to 37.12.","key_machinery":"The central object is the bi-level synergistic integration block, a fusion module that pools the concatenated modality features, computes softmax-normalized modality weights through two linear layers with ReLU, computes sigmoid-normalized channel weights in parallel, and then recalibrates each modality's feature map by multiplying the modality weight with the channel-weighted features. The other load-bearing components are the specific modality Mamba encoder, which runs each 3D modality through its own Mamba block, defined as a state-space sequence model with linear-time long-range modeling and forward/backward selective scans, followed by a residual block, and the decoder, which adds a Mamba block before upsampling and merges multi-scale encoder features through skip connections. The mechanism that carries the argument is the adaptive weighting itself: instead of fixed early fusion or concatenation, pooled descriptors drive dynamic emphasis of informative modalities and channels, which is what makes the fusion claim concrete.","core_discovery":"The central claim is that the proposed network, built from modality-specific 3D Mamba encoders, a bi-level synergistic integration block, and a decoder that restores spatial detail through skip connections, achieves the strongest reported tumor segmentation among the compared methods. The authors state that their approach outperforms all segmentation approaches in the BraTS2023 comparison, and that PET/CT fusion outperforms PET-only input on Hecktor2022. The design rationale is that Mamba gives linear-complexity long-range modeling, each modality keeps its distinct appearance through a dedicated encoder, and the two-level attention weights let the fusion adapt to which modality and which channels matter at each feature level. The paper identifies the fusion mechanism as the key contributor: adding it to a Mamba encoder raises the mean Dice score from 90.13 to 92.15 and cuts the mean Hausdorff distance from 3.78 to 2.62 mm.","pith_inferences":["Inference: because the BraTS2023 baseline numbers are copied from the SegMamba paper rather than re-run under this paper's own split and preprocessing, the true ranking under identical conditions is untested; a fair re-run could either confirm or narrow the reported gap.","Inference: the bi-level attention module is not tied to the specific Mamba block used here, so it could likely be attached to other state-space or Transformer encoders and still provide adaptive multimodal fusion.","Inference: the authors' stated future goal of quantifying per-modality contributions could be approximated immediately by ablating individual modality streams and recording Dice changes, which would indicate how much each MRI sequence or PET/CT stream contributes to the final segmentation.","Inference: with fixed hyperparameters of SGD, learning rate 1e-3, and 1000 epochs, the reported gains may depend on optimization settings; testing other optimizers and losses would clarify whether the advantage comes from the architecture or from favorable training conditions."],"forward_implications":["If the claim holds, per-modality Mamba encoders plus two-level attention becomes a reusable template for other aligned multimodal segmentation tasks, such as multi-contrast MRI or PET/CT in other body regions.","The reported gain of 2.02 Dice points from adding bi-level integration to the Mamba encoder suggests that the adaptive fusion mechanism contributes more than the choice of encoder backbone.","The strongest reported improvement is on enhancing tumor (ET), with a Dice score of 87.92 and a Hausdorff distance of 3.30 mm, which indicates the method particularly helps small, heterogeneous, boundary-ambiguous regions.","The Hecktor2022 comparison, showing 37.12 Dice for PET/CT fusion versus 33.07 for PET-only, supports the claim that combining metabolic and anatomical information is genuinely useful, not just adding more input channels."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space model that the modality-specific encoders are built on.","marker":"[3]"},{"why":"Defines the SegMamba baseline and is the stated source of all BraTS2023 comparison metric values in Table 2.","marker":"[19]"},{"why":"Swin UNETR, a Transformer baseline that the method claims to outperform on BraTS2023.","marker":"[5]"},{"why":"UNETR, a Transformer baseline that the method claims to outperform on BraTS2023.","marker":"[6]"},{"why":"SwinUNETR-V2, a Transformer baseline that the method claims to outperform on BraTS2023.","marker":"[7]"},{"why":"UX-Net, a CNN baseline that the method claims to outperform on BraTS2023.","marker":"[12]"},{"why":"SegresNet, a CNN baseline that the method claims to outperform on BraTS2023.","marker":"[14]"},{"why":"MedNeXt, a CNN baseline that the method claims to outperform on BraTS2023.","marker":"[16]"}],"fun_headline_variants":["Mamba fusion achieves SOTA 92.15 Dice in 3D tumor segmentation","Bi-level fusion lifts BraTS2023 Dice to 92.15","Mamba encoders plus adaptive fusion beat CNN and Transformer baselines","Bi-level attention fusion boosts Mamba 3D tumor segmentation","Linear-complexity Mamba model sets new 3D tumor segmentation record"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The baseline results on BraTS2023 are taken from the SegMamba paper rather than being re-run under this paper's own 70/10/20 split and preprocessing, so the reported advantage rests on the assumption that those published numbers are directly comparable.","fun_headline_variants_meta":{"raw":{"variants":["Mamba fusion achieves SOTA 92.15 Dice in 3D tumor segmentation","Bi-level fusion lifts BraTS2023 Dice to 92.15","Mamba encoders plus adaptive fusion beat CNN and Transformer baselines","Bi-level attention fusion boosts Mamba 3D tumor segmentation","Linear-complexity Mamba model sets new 3D tumor segmentation record"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00168,"raw_usage":{"total_tokens":6676,"prompt_tokens":974,"completion_tokens":5702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":5605}},"tokens_in":590,"tokens_out":5702,"duration_ms":39574,"temperature":1.0,"reasoning_tokens":5605,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:07:27.728945+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the seven compared methods under the exact same BraTS2023 split and preprocessing used in this paper, then test the mean Dice difference between this method and SegMamba, 92.15 versus 91.32, with a paired test across test cases; if the difference is not significant or reverses, the central outperformance claim fails.","supporting_citations":[{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI","cited_arxiv_id":null,"evidence_quote":"Defines the SegMamba baseline and is the stated source of all BraTS2023 comparison metric values in Table 2."},{"cited_title":"In: International MICCAI brainlesion workshop","cited_arxiv_id":null,"evidence_quote":"Swin UNETR, a Transformer baseline that the method claims to outperform on BraTS2023."},{"cited_title":"In: Proceedings of the IEEE/CVF winter conference on applications of computer vi- sion, WACV","cited_arxiv_id":null,"evidence_quote":"UNETR, a Transformer baseline that the method claims to outperform on BraTS2023."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI","cited_arxiv_id":null,"evidence_quote":"SwinUNETR-V2, a Transformer baseline that the method claims to outperform on BraTS2023."},{"cited_title":"In: The Eleventh International Conference on Learning Representations, ICLR (2023)","cited_arxiv_id":null,"evidence_quote":"UX-Net, a CNN baseline that the method claims to outperform on BraTS2023."},{"cited_title":"In: International MICCAI brainlesion workshop","cited_arxiv_id":null,"evidence_quote":"SegresNet, a CNN baseline that the method claims to outperform on BraTS2023."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI","cited_arxiv_id":null,"evidence_quote":"MedNeXt, a CNN baseline that the method claims to outperform on BraTS2023."}],"review_version":1}