{"id":"c6b1e417-e614-4bc8-ad81-3c04c6aff50a","arxiv_id":"2607.22718","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DAMamba-UNet3D combines encoder-only tri-plane Dynamic Adaptive Scan with a convolutional U-Net, reaching 0.815 mean Dice on BraTS 2020 at 5.3M parameters.","lead":"A new 3D U-Net design adds a learnable 'dynamic adaptive scan' to a Mamba state-space model, aiming to reach the segmentation quality of a larger model with far fewer parameters. The paper reports near-parity with SegMamba on BraTS 2020 at roughly 13x fewer parameters, and a larger variant that edges ahead by 0.5 Dice.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scan-pattern ablation is confounded: the only same-architecture fixed-vs-DAS comparison (both with bottleneck) is a tie (0.808 vs 0.807), while the reported DAS gain compares a bottleneck-free model to a bottleneck-containing baseline.","rationale":"The reader's weakest_assumption was that DAS might be inert and that no fixed-scan comparison exists inside the same architecture. Table 4 does contain a fixed-scan comparison, but it is confounded by the bottleneck placement: the fixed-scan rows use E2–E4+bottleneck, while the final DAS row removes the bottleneck. Thus the reader's concern is valid but under-specified: the problem is not that no comparison exists, but that the comparison that would isolate DAS (same architecture, simple scan swap) shows no benefit. This is a stronger, more concrete objection because it uses the paper's own numbers. The paper's self-acknowledged limitation in the Discussion and the lack of significance testing reinforce the need for conditional acceptance rather than rejection: the architecture and parameter-efficiency results are still useful, but the central DAS-vs-ToM claim is unsubstantiated. A single clean scan-swap experiment would settle the issue, so the conditional verdict remains appropriate.","tokens_in":6854,"tokens_out":6024,"duration_ms":50475,"concrete_test":"Run the final E2–E4-only architecture (convolutional bottleneck, no SSM at bottleneck) with fixed tri-orientated flattening replacing DAS, keeping all other settings identical (loss, training schedule, folds, full-volume evaluation). Compare the 5-fold mean Dice against the reported 0.815±0.013 for the DAS version, and compute per-fold paired differences (e.g., Wilcoxon signed-rank test across 5 folds). If the fixed-scan version reaches or exceeds 0.815, the DAS-specific advantage disappears. Additionally, run a matched DAS-vs-ToM comparison with the bottleneck block present, to confirm whether Table 4's apparent tie (0.807 vs 0.808) reproduces.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that learned tri-plane DAS appears competitive with and may improve upon fixed ToM scanning. The only experimental evidence intended to isolate scan type is Table 4. However, the caption states: 'Fixed orders use auxiliary E2–E4 + bottleneck layout.' Thus the fixed-scan rows (raster 0.805, snake 0.807, cross 0.806, tri-orientated 0.808) all include a bottleneck DAMamba block, while the final learned-DAS row (0.815) removes the bottleneck. The body text's comparison of 'learned DAS E2–E4 only (0.815)' against 'tri-orientated (0.808)' therefore conflates scan type with the removal of a bottleneck block. The matched comparison—learned DAS with bottleneck (0.807) vs. tri-orientated with bottleneck (0.808)—shows no improvement from DAS. The Discussion even concedes 'We do not isolate scan protocol from architecture in a single swap experiment.' At large scale, DAMamba-L vs. SegMamba differs in U-Net depth, block stacking, and bottleneck design, which the paper admits prevents attributing the gain to scan type. Consequently, the evidence does not establish that DAS outperforms ToM; differences of 0.007 Dice are also within the reported standard deviations, and no significance testing is provided. This is a load-bearing gap because the paper's central claim and novelty hinge on DAS being beneficial in a hybrid U-Net, not merely on the overall architecture.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DAMamba-UNet3D, a hybrid 3D U-Net in which encoder stages E2–E4 contain tri-plane Dynamic Adaptive Scan (DAS) Mamba blocks and convolutions are retained elsewhere. On BraTS 2020 five-fold full-volume evaluation, the compact model (~5.3M parameters) attains 0.815±0.013 mean Dice versus 0.824±0.014 for retrained SegMamba (~70M); a large variant DAMamba-L (~70M) reaches 0.829±0.012, 0.5 pt higher than SegMamba. The authors claim the first medical-segmentation DAS architecture and the first 3D extension of DAS, and argue that learned tri-plane DAS is competitive with and may improve upon fixed Tri-orientated Mamba (ToM) scanning.","tokens_in":7239,"tokens_out":3994,"duration_ms":37203,"significance":"If the central claim were established, this work would provide a parameter-efficient alternative to fixed-scan Mamba segmenters and a meaningful 3D extension of an adaptive scan operator. The paper has notable strengths: it uses a standard public benchmark (BraTS 2020), a clearly described five-fold full-volume evaluation protocol, and transparent reporting of training settings. However, the evidence for the central DAS-versus-ToM claim is confounded by the scan-pattern ablation design, and the paper itself concedes that scan protocol and architecture are not isolated. As presented, the significance of the claimed adaptive-scan advantage is not yet demonstrated; what is demonstrated is that a particular hybrid U-Net with encoder-only DAS blocks is competitive at low parameter count.","major_comments":[{"comment":"The scan-pattern ablation conflates scan type with bottleneck removal. The caption states that fixed orders use 'auxiliary E2–E4 + bottleneck layout,' while the final learned-DAS row removes the bottleneck. The only matched comparison — learned DAS with bottleneck (0.807) versus tri-orientated with bottleneck (0.808) — shows no gain from DAS. The +0.7 pt difference between learned DAS (final, 0.815) and tri-orientated (0.808) therefore cannot be attributed to DAS; it is attributable to removing the bottleneck block. The Discussion admits this ('We do not isolate scan protocol from architecture in a single swap experiment'), but the Abstract and Conclusion assert a stronger claim. Please supply a same-layout fixed-versus-DAS comparison, e.g., fixed scans with E2–E4 only, and report paired significance tests.","section":"§3.2, Table 4"},{"comment":"The sentence 'learned DAS ahead of fixed orders when encoder placement is matched' is directly contradicted by Table 4's matched row: learned DAS with bottleneck is 0.807, tri-orientated with bottleneck is 0.808. Either the claim must be corrected to state that DAS is ahead only when the bottleneck is removed, or new matched experiments are needed. This is a report-of-results error, not merely an interpretation issue.","section":"§3.2 text"},{"comment":"The claim that DAMamba-L 'surpasses SegMamba ... suggesting that learned DAS in an encoder-only hybrid design can outperform fixed ToM' is unsupported as stated. DAMamba-L differs from SegMamba in U-Net depth, block stacking, and bottleneck design, as the authors acknowledge. The improvement, even if real, could come entirely from those architectural differences. To support the attribution to DAS, the paper needs either a scale-matched ablation that varies only the scan mechanism, or a much more cautious conclusion limited to the full DAS-native hybrid architecture.","section":"§4, Discussion and Table 2"},{"comment":"The key numerical comparisons (0.815 vs 0.824, 0.829 vs 0.824, 0.807 vs 0.808, 0.808 vs 0.815) are all within one standard deviation or less, and no significance testing is reported. Since the protocol uses five subject-level folds, paired tests (e.g., Wilcoxon or paired t on fold Dice) are straightforward and would materially strengthen the claims. The Limitations section correctly notes that statistical testing is needed, but the abstract and conclusion nevertheless draw positive conclusions without it.","section":"§2.6, §3, Table 3/4"}],"minor_comments":[{"comment":"The abstract provides a GitHub code URL, while the Conclusion states 'We will publish our code upon acceptance of this paper.' Please reconcile these statements and make the availability status unambiguous.","section":"Abstract vs. Conclusion"},{"comment":"The tri-plane DAS operation is defined only as an average of 2D DAS on axial, coronal, and sagittal planes. It would clarify whether the offsets are shared or separately learned for each plane, and how the 2D operator is adapted when the plane dimensions differ (e.g., D×W for coronal). This affects the reader's ability to reproduce the method.","section":"Eq. (4)"},{"comment":"The loss equation uses per-class weights w_i^(c) but does not define the class index c or the range of i. A short definition would help reproducibility.","section":"Eq. (5)"},{"comment":"Reference [7] lists an incomplete author list ('Zhou, Y., Zhang, Y., et al.'), and reference [8] contains a broken URL fragment ('https–www'). Please fix these formatting issues.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is borderline. The 3D DAS idea is novel and the evaluation protocol is careful, but the central claim of DAS superiority over fixed ToM is not supported by the current experimental design because the only matched comparison is a tie and the apparently positive comparison removes a bottleneck block. The authors could strengthen the paper with a matched scan-pattern ablation (preferably at both compact and large scale) and significance tests. The code-availability inconsistency should also be fixed. Scope-wise, the work fits a medical-imaging venue interested in efficient Mamba segmenters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful part of this paper is the engineering: it is the first 3D medical use of dynamic adaptive scan (DAS), it formulates DAS in a tri-plane way, and it reports an encoder-only placement rule that seems to help. The compact model gets 0.815 Dice at 5.3M parameters versus 0.824 for the retrained 70M SegMamba, under the same folds, loss, and full-volume evaluation. That near-parity at 13x fewer parameters is a legitimate empirical finding, and the protocol is described clearly enough to reproduce once code is out.\n\nThe soft spot is the central claim. The paper says learned DAS may improve over fixed Tri-orientated Mamba, but Table 4 does not actually test that. The fixed-order rows all include a bottleneck DAMamba block; the final learned-DAS row removes the bottleneck. So the 0.815 versus 0.808 comparison conflates scan type with architecture. The one matched comparison—learned DAS with bottleneck versus tri-orientated with bottleneck—is a tie (0.807 vs 0.808). The Discussion admits this, and the limitations section concedes there is no multi-dataset validation or statistical testing. The 0.5-point gain of DAMamba-L over SegMamba is also within one standard deviation, so without a swap experiment or significance testing, the DAS-specific benefit is unproven.\n\nI would not call this a fatal flaw, because the paper is honest about the limitation and the near-parity result stands on its own. But the framing oversells DAS. The paper would be much stronger if the authors either provided a same-architecture scan-swap experiment (only the scan type varies) or explicitly demoted DAS to a component that does not clearly outperform fixed scans.\n\nWho should read it: people working on efficient SSM segmenters, and anyone who wants a cautionary example of a confounded ablation in medical imaging. I'd bring it to a reading group as a methods critique, but I would not cite it for the DAS claim until the confound is resolved.\n\nRecommendation: send it to peer review, but require the scan-swap experiment or a revised claim. The paper's evidence will support a modest claim about parameter efficiency; it does not yet support the DAS superiority claim.","headline":"A cleanly reported architecture paper whose central DAS-vs-ToM claim is not actually isolated—but the 13x parameter reduction at near-parity is real and worth a referee's time.","tokens_in":7701,"tokens_out":1507,"would_cite":false,"duration_ms":16013,"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":"This paper claims that learned, data-dependent scan ordering — Dynamic Adaptive Scan in a 3D Mamba U-Net — matches a fixed tri-orientated Mamba segmenter at 13× lower parameter count and beats it when scaled up.","keywords":["dynamic adaptive scan","state space model","Mamba","3D medical image segmentation","brain tumor segmentation","parameter efficiency","U-Net","BraTS 2020"],"falsifier":"Train the identical DAMamba-UNet3D architecture but replace the tri-plane DAS with a fixed tri-orientated flattening (same placement, same parameter count, same pipeline). If mean Dice stays at 0.815 or the gap to the baseline disappears, the adaptive mechanism is not responsible for the result. Alternatively, dump the learned offsets on a few test volumes and check whether they systematically deviate from the identity grid and focus on tumor boundaries.","tokens_in":6729,"feed_emoji":"🧠","tokens_out":5623,"duration_ms":51278,"temperature":0.7,"pith_summary":"This paper tries to show that a U-Net for 3D medical image segmentation can get the long-range context benefits of Mamba-style state-space models without paying for enormous parameter budgets, by letting the model learn the order in which it scans the volume. The authors build a hybrid U-Net where only the encoder's middle three stages contain Mamba blocks that reorder voxels using learned, data-dependent offsets (Dynamic Adaptive Scan) applied across three anatomical planes, while the rest of the network stays convolutional. On BraTS 2020 brain tumor segmentation, the compact version (5.3M parameters) nearly matches a fixed-scan Mamba segmenter of about 70M parameters, and a wider version surpasses it. The central claim is that adaptive scan ordering is a viable, parameter-efficient alternative to hand-designed fixed scan patterns for volumetric segmentation.","feed_headline":"13x smaller model matches fixed-scan model in brain tumor segmentation","feed_subtitle":"Learned, data-driven scan order lets a compact U-Net match a 70M-parameter baseline on BraTS 2020 tumor segmentation.","key_machinery":"The central mechanism is the tri-plane Dynamic Adaptive Scan (DAS3D) block. DAS predicts 2D offsets per position in each of the three anatomical planes (axial, coronal, sagittal), samples features with bilinear interpolation, averages the three reordered maps, and then runs the standard selective scan of the state-space model along the resulting 1D sequence. This turns the scan trajectory from a fixed design choice into a data-dependent operation that costs O(N) time per block, preserving the efficiency of Mamba while adding adaptability. An important secondary finding is that placing these blocks only in the encoder stages E2–E4, with convolutions at the bottleneck and decoder, is critical","core_discovery":"The paper's central discovery is that Dynamic Adaptive Scan — a learning-based reordering of the voxel sequence fed to a selective state-space model — can be lifted to 3D by averaging scans over axial, coronal, and sagittal planes, and that this tri-plane formulation works well enough to make a 5.3M-parameter U-Net competitive with a 70M-parameter fixed-scan Mamba segmenter (0.815 vs 0.824 mean Dice on BraTS 2020). With a wider variant matched to the baseline's scale, the adaptive-scan model exceeds the fixed-scan model (0.829 vs 0.824). The authors present this as the first application of DAS to medical segmentation and the first 3D extension of DAS, and they argue that scan order, not just","pith_inferences":["Because the paper never inspects the learned offsets, the 'adaptive' part may be contributing less than the architectural differences (block placement, width, stacking); a direct swap of DAS for a fixed scan inside the identical model would isolate the effect.","The tri-plane averaging is a simple 2D-per-plane lifting; a genuinely 3D DAS (e.g., volumetric deformable sampling) might show larger gains, especially for non-axis-aligned tumor geometry.","If the adaptive ordering is truly learning meaningful tumor-first scanning, the learned offsets should be visualizable as concentration near lesion boundaries; that is a testable prediction of the paper's mechanism."],"forward_implications":["Near-parity with a 13× larger fixed-scan model shows that global context can be bought with learned scan order rather than parameter count.","The encoder-only placement principle — no SSM blocks at bottleneck or decoder — gives a concrete design rule for hybrid U-Nets.","DAMamba-L's +0.5 pt Dice gain over the fixed-scan baseline suggests learned scanning can outperform fixed multi-orientation flattening when capacity is matched.","The O(N) cost of the adaptive scan preserves Mamba's linear-time advantage, making the approach usable on high-resolution 3D volumes."],"fun_headline_variants":["Adaptive scan U-Net: 13x fewer params, near-par tumor segmentation","5.3M-param Mamba U-Net rivals 70M fixed-scan on BraTS","Learning scan order: small 3D U-Net matches big Mamba model","DAMamba-UNet3D: dynamic scan beats fixed Mamba at 1/13 size","First 3D dynamic scan for medical segmentation, tiny and competitive"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that averaging three 2D adaptive scans produces a meaningful 3D scan order; the paper offers no evidence about the learned offsets, so if they collapse to identity or random noise the adaptive mechanism is inert and the reported differences come from other architectural choices.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive scan U-Net: 13x fewer params, near-par tumor segmentation","5.3M-param Mamba U-Net rivals 70M fixed-scan on BraTS","Learning scan order: small 3D U-Net matches big Mamba model","DAMamba-UNet3D: dynamic scan beats fixed Mamba at 1/13 size","First 3D dynamic scan for medical segmentation, tiny and competitive"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1354,"prompt_tokens":938,"completion_tokens":416,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":312}},"tokens_in":682,"tokens_out":416,"duration_ms":4798,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:52:35.514610+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical DAMamba-UNet3D architecture but replace the tri-plane DAS with a fixed tri-orientated flattening (same placement, same parameter count, same pipeline). If mean Dice stays at 0.815 or the gap to the baseline disappears, the adaptive mechanism is not responsible for the result. Alternatively, dump the learned offsets on a few test volumes and check whether they systematically deviate from the identity grid and focus on tumor boundaries.","supporting_citations":[],"review_version":1}