{"id":"69b3f844-df05-4fe8-baae-632c2b405e8c","arxiv_id":"2501.00751","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HCMA-UNet reports state-of-the-art Dice scores on three DCE-MRI breast cancer datasets using a compact CNN-Mamba UNet with axial self-attention and a feature-guided region-aware loss.","lead":"This paper presents a neural network, HCMA-UNet, that segments breast cancer lesions in contrast-enhanced MRI scans. It combines lightweight convolutions, Mamba sequence modeling, and axial attention to reach high accuracy with only 2.87 million parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The state-of-the-art claim rests on a single fixed training protocol applied to all baselines; without per-model tuning, repeated runs, or statistical tests, the Table I gaps could be configuration artifacts.","rationale":"I read the paper in good faith and find no internal mathematical inconsistency in the proposed architecture or loss. The ablations in Tables II and III are monotonic and plausible, the parameter counts are self-consistent, and the qualitative comparison supports the design choices. The decisive weakness is the external validity of the SOTA claim: all baselines are trained under a single protocol not tailored to each model, with no repeated runs, error bars, or statistical testing. The reader's weakest_assumption identifies exactly this baseline-tuning issue, and I agree that it is load-bearing because the entire contribution is framed as 'state-of-the-art performance with reduced computational complexity.' If the baselines are under-tuned, the headline result may be an artifact of the comparison protocol rather than the architecture. This is an addressable experimental-design problem, not a reason to reject the work outright, so the conditional verdict remains appropriate. No code or artifacts were available to inspect, which further justifies caution but does not by itself change the verdict.","tokens_in":9379,"tokens_out":9337,"duration_ms":93173,"concrete_test":"Re-run the comparison on Dataset I using each baseline's official or recommended training configuration (e.g., nnU-Net with its self-configuring defaults, SwinUNETR and U-Mamba with their original hyperparameters), while keeping the same training/test split and the same FRLoss integration for the w/ FRLoss rows. Run at least three seeds per method, report mean ± std Dice, and perform paired significance tests against HCMA-UNet. If any baseline matches or exceeds HCMA-UNet's 80.28 Dice, or if the differences are not statistically significant, the state-of-the-art claim is not supported. A cheaper preliminary check is to verify whether the Table I baseline Dice on the public MAMA-MIA dataset are consistent with published results for those architectures on the same benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that HCMA-UNet achieves state-of-the-art performance depends entirely on Table I, where nine baseline architectures are trained under one uniform protocol (patch 128^3, batch size 2, 500 epochs, AdamW lr 1e-4, Section III-C). This protocol is not equally appropriate for all baselines: SwinUNETR, UMamba, LKM-UNet, and MedNeXt have published training recipes with different patch sizes, learning-rate schedules, and epoch counts, and some are known to be sensitive to batch size and regularization. Forcing them into a single recipe can suppress their performance by amounts comparable to the reported gaps (HCMA-UNet leads the best baseline by 1.8–3.0 Dice on Dataset I and 0.7–3.2 Dice on Dataset III). The paper reports no validation split, no per-model hyperparameter search, no multiple-seed averaging, and no significance tests, so the observed gaps cannot yet be attributed to architectural superiority rather than configuration artifact. Because the SOTA claim is the paper's headline contribution, this is the most load-bearing weaker assumption. The architecture itself and the FRLoss ablations appear internally coherent; the issue is specifically that the comparison that justifies SOTA is not yet controlled.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HCMA-UNet, a hybrid CNN-Mamba UNet for 3D breast cancer lesion segmentation in DCE-MRI. The architecture combines a lightweight residual CNN backbone with a Multi-view Axial Self-Attention Mamba (MISM) module, which uses Visual State Space Blocks and axial self-attention under an asymmetric channel-splitting scheme to process the three anatomical views. The paper also introduces a Feature-guided Region-aware Loss (FRLoss) composed of positive compactness, boundary-aware, and hard-negative-mining terms. The method is evaluated on one private and two public datasets against nine baselines, reporting state-of-the-art Dice scores of 80.28%, 76.14%, and 78.05% with 2.87M parameters and 126.44 GFLOPs. Ablations in Tables II and III support the individual contributions of the MISM components and the FRLoss terms. Source code is made available.","tokens_in":9657,"tokens_out":8271,"duration_ms":81214,"significance":"If the reported results hold, the paper makes a useful practical contribution: a compact architecture that reportedly outperforms much larger models while using substantially fewer parameters and FLOPs, together with a loss function that improves several baseline architectures. The availability of code, the internal consistency of the reported numbers, and the component-wise ablations are strengths. However, the central state-of-the-art claim currently rests on a baseline comparison that is not sufficiently controlled, and the lack of uncertainty quantification leaves the size of the reported gains unverified. The architectural idea and the proposed loss are coherent and worth pursuing, but the evidence presented does not yet support the headline claim as stated.","major_comments":[{"comment":"The state-of-the-art claim rests on a comparison in which all nine baselines were trained under a single fixed protocol (patch size 128^3, batch size 2, 500 epochs, AdamW with learning rate 1e-4), with no per-method hyperparameter tuning, no multiple-seed averaging, and no statistical tests. Published recipes for several baselines (e.g., SwinUNETR, UMamba, LKM-UNet, MedNeXt) differ substantially in patch size, learning-rate schedule, and training length, and some are sensitive to batch size and regularization. The reported Dice gaps over the best baseline are 1.85, 2.30, and 2.52 points on Datasets I, II, and III respectively, which are within the range that configuration artifacts can produce. The authors should either tune each baseline under the same budget, report results from the baselines' own recommended settings, or explicitly restrict the claim to the fixed protocol and add significance testing.","section":"Section III-C, Table I"},{"comment":"Statistical robustness is not established. Dataset III has only 25 test samples from a random split, Dataset I's random split is not described as seeded, and no error bars, confidence intervals, or significance tests are reported anywhere. Run-to-run variability from weight initialization and data ordering can be comparable to the reported improvements in medical segmentation benchmarks, so the cross-dataset gains in Table I and the component gains in Tables II and III need to be supported by repeated runs, multiple random splits, or paired statistical tests. Without this, the claims of consistent and significant improvement are not quantitatively supported.","section":"Section III-A/B, Table I"},{"comment":"The paper states that the Axial Self-Attention mechanism 'achieves linear computational complexity.' The scaled dot-product attention in Eq. (2) computes pairwise similarities over the L3 dimension, which is O(H*W*L3^2*C) operations per volume, i.e., quadratic in the slice dimension and super-linear in the total volume H*W*L3. This is not linear complexity. If the intended meaning is linear in the slice length per query, that should be stated explicitly; otherwise the efficiency claim as written is inaccurate and should be corrected. The reported GFLOPs may still be favorable, but the complexity statement needs revision.","section":"Section II-B2, Eqs. (1)-(2)"}],"minor_comments":[{"comment":"There is a typo: 'surrocunding tissues' should be 'surrounding tissues.'","section":"Introduction, first paragraph"},{"comment":"The 50/25/25 channel split in ASC is justified by a heuristic about the axial plane displaying complete and symmetrical features, but no ablation with alternative split ratios is provided. A sensitivity analysis of this ratio would make the design choice more convincing.","section":"Section II-B3"},{"comment":"The symbol N is used both for the number of top hard-negative pixels and for the dilated hard-negative set obtained after the dilation operation. Rename one of the two to avoid notational confusion.","section":"Section II-D4, Eq. (7)"},{"comment":"The FRLoss hyperparameters (lambda=5, T1=10, T2=10, N=250) are fixed across all experiments without sensitivity analysis. Please report whether the results are stable around these values, particularly across the three datasets.","section":"Section III-C"},{"comment":"The single-slice Dice comparison in Fig. 4 (a 10.81% improvement over UMamba) is anecdotal. A quantitative summary over the test set or several representative slices would be more informative than one selected case.","section":"Section III-D, Fig. 4"},{"comment":"Please specify the input size and the exact definition of GFLOPs used for the complexity numbers, for example whether they correspond to a forward pass on a patch of size 128^3. This is needed for a fair comparison with the baseline complexity figures.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The main bottleneck is the uncontrolled baseline comparison in Table I; this is fixable and the paper is otherwise internally coherent. I see no grounds for rejection, but the state-of-the-art claim should not be accepted without per-method tuning or equivalent safeguards, repeated runs, and statistical tests."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely new combination of Mamba scanning, axial attention, and a channel-split trick, and the internal ablations are clean. The weak point is the baseline comparison: all nine competitors are forced into one training protocol, with no error bars or repeated runs, so the 'state-of-the-art' headline is not yet supported.\n\nWhat's new: the MISM module reslices 3D volumes into three planes, uses VSSB (from VMamba) for intra-slice scanning, adds one-dimensional axial self-attention for inter-slice relations, and splits channels asymmetrically (50/25/25) to cut redundancy. The FRLoss is a sensible composition of positive compactness, boundary-aware, and hard-negative mining losses. The ablations in Tables II and III give each component a small, consistent bump, which is credible. The efficiency figures are real: 2.87M parameters and 126 GFLOPs is an order of magnitude lighter than UMamba or nnU-Net, and the per-slice design is a reasonable way to keep 3D context without quadratic cost.\n\nThe soft spots are exactly where the stress-test note points. The uniform training protocol (patch 128^3, batch size 2, 500 epochs, AdamW lr 1e-4) is not appropriate for all baselines. nnU-Net, in particular, is a self-configuring framework; forcing it into a hand-set recipe undermines the comparison. Likewise SwinUNETR and LKM-UNet have published recipes that differ. No validation split is described, no multiple seeds, no significance tests, and Dataset III has only 25 test samples, so the 0.7–3.2 Dice gaps on that dataset could easily be noise. That said, the gaps are directionally consistent across all three datasets, and the architecture itself is plausible. I would not call the paper flawed; I would say the SOTA claim is unearned by the evidence as presented.\n\nThe FRLoss generalization result (w/ vs w/o rows for every baseline) is a nice touch, and those numbers also lack error bars, but the pattern is uniform. The code link is given, though nothing is verifiable from the text.\n\nWho this is for: anyone working on efficient 3D medical segmentation, or on loss designs for small lesions. It deserves a serious referee, but the referee should push for stronger empirical rigor: per-model hyperparameter selection or at least multiple seeds with means and standard deviations, plus a proper statistical test on the private dataset. A revision along those lines would make the contribution far more solid.\n\nRecommendation: send to peer review, with the expectation of major revision on the benchmarking side.","headline":"A coherent hybrid CNN-Mamba UNet with a clean internal ablation story, but the state-of-the-art claim rests on a fixed-protocol baseline comparison without error bars or significance tests.","tokens_in":10208,"tokens_out":2810,"would_cite":true,"duration_ms":26530,"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 3D hybrid CNN-Mamba UNet with axial self-attention reaches state-of-the-art Dice on three breast DCE-MRI datasets while using a fraction of the compute of prior Mamba models.","keywords":["breast cancer segmentation","DCE-MRI","Mamba","axial self-attention","medical image segmentation","hybrid CNN-Mamba","FRLoss"],"falsifier":"Retrain the nine baselines on the same three datasets using each method's own recommended or individually tuned settings and several random seeds, then compare Dice distributions. If a tuned baseline matches or exceeds HCMA-UNet's Dice, or if HCMA-UNet's lead falls inside run-to-run variation, the state-of-the-art claim is refuted; a second check is to remove FRLoss and confirm the paper's reported 1.01-point Dice improvement disappears.","tokens_in":9155,"feed_emoji":"🩺","tokens_out":14045,"duration_ms":109498,"temperature":0.7,"pith_summary":"The paper sets out to establish that a compact hybrid architecture can segment breast cancer lesions in DCE-MRI more accurately and far more cheaply than existing CNN, Transformer, and Mamba models. Its central claim is that HCMA-UNet, a lightweight CNN backbone combined with a Multi-view Axial Self-Attention Mamba (MISM) module, reaches Dice scores of 80.28%, 76.14%, and 78.05% on one private and two public datasets while using only 2.87M parameters and 126.44 GFLOPs. The paper also proposes a Feature-guided Region-aware Loss (FRLoss) that adds roughly one Dice point to HCMA-UNet and, according to its comparison tables, improves most of the nine baseline models when swapped in. A sympathetic reader would take the contribution to be a practical recipe: replace costly whole-volume Mamba scanning with tri-planar slicing plus axial self-attention, and guide training with a feature-similarity loss, to get accurate segmentation at a fraction of the compute.","feed_headline":"A 2.87M-parameter hybrid UNet beats nine MRI rivals","feed_subtitle":"Hybrid CNN-Mamba with axial self-attention hits 80.28% Dice at a fraction of previous Mamba compute","key_machinery":"The load-bearing object is the Multi-view Axial Self-Attention Mamba (MISM) module, a hybrid block that replaces direct 3D Mamba scanning with tri-planar processing: the volume is cut into axial, coronal, and sagittal slices, each slice runs through VMamba's Visual State Space Block (VSSB) with bidirectional cross-scanning for intra-slice features, and a linear-complexity 1D Axial Self-Attention (ASA) mechanism then models inter-slice correlations along the orthogonal direction. The Asymmetric Split Channel (ASC) strategy feeds 50% of the channels through the axial plane and 25% through each of the other planes, which the paper argues cuts redundancy because the axial plane already shows the complete structure. The second carrier is the Feature-guided Region-aware Loss (FRLoss), computed on feature maps before the output block: it averages foreground features into a foreground center, pulls positive pixels toward that center (positive compactness), penalizes negative pixels near the boundary that resemble the center (boundary aware), and mines the top-N most similar negative pixels (hard negative mining).","core_discovery":"On the paper's own terms, HCMA-UNet is a 3D encoder-decoder whose encoder combines a MobileNetV2-style Res Block and a Hybrid Block, and whose decoder uses transpose convolutions with HCMA-Up blocks. The novel part is the MISM module, which reslices each 3D volume along the axial, coronal, and sagittal planes instead of flattening the whole volume into one sequence; each plane is processed by VMamba's Visual State Space Block for intra-slice context, and an Axial Self-Attention mechanism runs along the orthogonal direction to model inter-slice dependencies. An Asymmetric Split Channel strategy allocates 50% of the channels to the axial view and 25% to each of the other two, reducing redundant computation. With this design the paper reports state-of-the-art Dice on all three datasets, outperforming nine baselines including nnU-Net, SwinUNETR, UNETR++, U-Mamba, LKM-UNet, and MambaClinix, and the FRLoss experiments show consistent gains when its three terms are added one by one.","pith_inferences":["A testable extension the paper leaves implicit: the 50/25/25 axial-first channel split is motivated by the axial plane showing the complete breast structure, so a learnable or dataset-adaptive split could be benchmarked in organs without a dominant axis, such as liver or brain.","Because FRLoss attacks a generic failure mode—negative pixels whose features resemble the foreground center—it plausibly transfers to other modalities and lesion types, but the paper demonstrates it only on DCE-MRI breast data, so that transfer is my inference rather than a paper claim.","The shared training protocol makes the headline comparison a configuration-equality test rather than a best-configuration test; an external benchmark with per-method tuned settings and multi-seed confidence intervals would separate genuine architectural gains from configuration effects.","If FRLoss's cross-architecture gains are real, a lighter design point follows that the paper does not explore: keep a fixed backbone and add FRLoss, since the loss terms cost nothing at inference time."],"forward_implications":["If the reported results hold, 3D breast lesion segmentation no longer requires multi-hundred-million-parameter Mamba stacks: HCMA-UNet uses roughly 4% of U-Mamba's parameters and about 1% of its FLOPs while scoring higher Dice on all three datasets.","The tri-planar MISM recipe gives a general way to adapt sequence-native Mamba blocks to volumetric medical images without flattening the volume into an unnatural 1D order, so the approach should transfer to other 3D segmentation tasks.","FRLoss behaves as a plug-in loss: the comparison tables show that most of the nine baselines improve when FRLoss is added, which means existing DCE-MRI segmentation models can likely be upgraded without changing their architectures.","On the private and public datasets evaluated, the complete HCMA-UNet with FRLoss sets a concrete numeric benchmark (Dice 80.28 / 76.14 / 78.05) for future breast cancer segmentation work to compare against."],"supporting_citations":[{"why":"Supplies the selective state-space sequence model whose unidirectional scan limitation MISM is designed to overcome.","marker":"[9]"},{"why":"Provides the Visual State Space Block (VSSB) and cross-scan bidirectional mechanism reused for intra-slice feature extraction.","marker":"[13]"},{"why":"Defines the nnU-Net preprocessing, augmentation, and patch-based training protocol used for all experiments, and is a CNN baseline the paper must outperform.","marker":"[18]"},{"why":"U-Mamba is the main Mamba-based 3D segmentation baseline; HCMA-UNet compares against it at far lower parameter and FLOP counts.","marker":"[10]"},{"why":"LKM-UNet is a Mamba-UNet baseline in Table I used for comparison.","marker":"[11]"},{"why":"MambaClinix is a Mamba-based baseline whose Dataset I Dice is among the highest the paper has to beat.","marker":"[12]"},{"why":"SwinUNETR is a Transformer baseline in Table I that the state-of-the-art claim must beat.","marker":"[4]"},{"why":"UNETR++ is a Transformer baseline in Table I whose scores anchor the comparison.","marker":"[5]"},{"why":"MAMA-MIA is public Dataset II used for evaluation and comparison.","marker":"[21]"},{"why":"Yunnan Cancer Hospital public DCE-MRI dataset is Dataset III used for external evaluation.","marker":"[22]"}],"fun_headline_variants":["Hybrid CNN-Mamba UNet hits SOTA with 2.87M params","Axial self-attention Mamba UNet slashes MRI segmentation compute","2.87M-param CNN-Mamba beats nnU-Net and 8 other MRI models","Tiny UNet with axial attention triples efficiency, keeps SOTA Dice","Lightweight CNN-Mamba UNet: SOTA breast MRI segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes every baseline model performs near its best under the single shared training recipe (patch size $128^3$, batch size 2, 500 epochs, AdamW at $10^{-4}$), with no method-specific tuning or multi-seed averaging, so the reported gaps could be configuration artifacts rather than real architectural differences.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid CNN-Mamba UNet hits SOTA with 2.87M params","Axial self-attention Mamba UNet slashes MRI segmentation compute","2.87M-param CNN-Mamba beats nnU-Net and 8 other MRI models","Tiny UNet with axial attention triples efficiency, keeps SOTA Dice","Lightweight CNN-Mamba UNet: SOTA breast MRI segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3519,"prompt_tokens":991,"completion_tokens":2528,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2437}},"tokens_in":607,"tokens_out":2528,"duration_ms":17165,"temperature":1.0,"reasoning_tokens":2437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:42:38.933570+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the nine baselines on the same three datasets using each method's own recommended or individually tuned settings and several random seeds, then compare Dice distributions. If a tuned baseline matches or exceeds HCMA-UNet's Dice, or if HCMA-UNet's lead falls inside run-to-run variation, the state-of-the-art claim is refuted; a second check is to remove FRLoss and confirm the paper's reported 1.01-point Dice improvement disappears.","supporting_citations":[{"cited_title":"Vmamba: Visual state space model,","cited_arxiv_id":null,"evidence_quote":"Provides the Visual State Space Block (VSSB) and cross-scan bidirectional mechanism reused for intra-slice feature extraction."},{"cited_title":"nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Defines the nnU-Net preprocessing, augmentation, and patch-based training protocol used for all experiments, and is a CNN baseline the paper must outperform."},{"cited_title":"Lkm-unet: Large kernel vision mamba unet for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"LKM-UNet is a Mamba-UNet baseline in Table I used for comparison."},{"cited_title":"Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,","cited_arxiv_id":null,"evidence_quote":"SwinUNETR is a Transformer baseline in Table I that the state-of-the-art claim must beat."},{"cited_title":"Unetr++: delving into efficient and accurate 3d medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"UNETR++ is a Transformer baseline in Table I whose scores anchor the comparison."},{"cited_title":"A robust and efficient ai assistant for breast tumor segmentation from dce-mri via a spatial-temporal framework,","cited_arxiv_id":null,"evidence_quote":"Yunnan Cancer Hospital public DCE-MRI dataset is Dataset III used for external evaluation."}],"review_version":1}