{"id":"2e32a6d5-a18e-4abb-9e84-3df6748382f7","arxiv_id":"2501.08163","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DH-Mamba is a dual-domain hierarchical Mamba network that uses circular k-space scanning and local diversity enhancement to outperform prior MRI reconstruction methods on three public datasets.","lead":"This paper introduces DH-Mamba, a deep learning model for accelerated MRI reconstruction that applies Mamba state space models in both image and k-space domains with a circular scanning strategy. The method aims to achieve fast, high-quality reconstruction with lower computational cost than existing approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III contains an impossible NMSE value for DH-Mamba at AF=8 on SKM-TEA; this internal inconsistency is a load-bearing flaw in the multi-coil claim until corrected.","rationale":"The reader correctly flagged Table III as a serious reporting error, and my independent reading agrees that it is the strongest concrete defect. However, I go further: the specific NMSE value 0.0020 ± 0.0032 is not merely anomalous; it is statistically impossible, and it directly affects the claimed 'consistently surpasses' result on the multi-coil SKM-TEA dataset. The reader's own weakest assumption concerned the circular k-space scan being a heuristic without proof. I do not regard that as the most load-bearing concern because the ablation in Table VI provides empirical support for its contribution, even if the design rationale is informal. The impossible NMSE, by contrast, is an internal inconsistency in the headline evidence that cannot be resolved from the manuscript alone. The recommended verdict is unchanged: the paper should remain conditional on correction of Table III and on public reproduction of the multi-coil numbers.","tokens_in":21209,"tokens_out":3988,"duration_ms":40801,"concrete_test":"Reproduce the SKM-TEA AF=8 evaluation using the authors' released code and the exact test split and mask seed described in Section IV-A. Compute DH-Mamba's NMSE, SSIM, and PSNR over the same 21 test volumes. If the corrected NMSE at AF=8 is around 0.020 or above (closer to or above ReconFormer's 0.0239), the Table III entry is wrong and the multi-coil claim needs revision. If the corrected NMSE is below 0.005, the discrepancy is a typo in the reported mean or error bar and the central multi-coil claim survives. As a secondary check, recompute Tables I and II from the same pipeline to confirm no similar order-of-magnitude transcription errors exist elsewhere.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DH-Mamba 'consistently surpasses' prior methods on all three datasets depends on the correctness of the quantitative tables. In Table III (SKM-TEA, multi-coil), the reported NMSE for DH-Mamba at AF=8 is 0.0020 ± 0.0032. This is internally impossible: NMSE is non-negative, so a mean of 0.0020 with a standard deviation of 0.0032 implies a negative lower tail, and it would mean that 8x undersampling produces roughly eight times lower NMSE than 4x undersampling (0.0156) despite a lower PSNR (32.97 vs. 35.43). The entry is almost certainly a transcription error, e.g., the intended value may be 0.0202 or 0.0220. If corrected to a plausible value around 0.020, the AF=8 advantage over ReconFormer (0.0239) shrinks substantially, weakening the claim of consistent superiority on the multi-coil dataset. Because the paper only promises future code release and provides no reproducibility appendix, this single impossible number cannot be resolved without external verification. This is more load-bearing than the heuristic nature of the circular k-space scan, since the ablation in Table VI already demonstrates a real 0.56 dB contribution from that component.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DH-Mamba, a dual-domain hierarchical state space model for accelerated MRI reconstruction. The method combines a k-space branch with a circular scanning strategy designed to respect the concentric frequency structure of k-space, an image-space branch using standard Mamba scanning, a hierarchical scanning strategy that processes one high-resolution path and three downsampled paths to reduce sequence length, and a local enhancement module (LEM) that applies a pixel-wise gating mask to increase local feature diversity. The architecture is evaluated on three public datasets (CC359, fastMRI, SKM-TEA) under Cartesian, radial, and random undersampling masks at acceleration factors 4 and 8 (and additional factors in the mask experiments). The central claim is that DH-Mamba consistently outperforms prior CNN-, transformer-, and Mamba-based methods while using lower computation (117 GFLOPs versus 342 G for ReconFormer and 190 G for MambaIR).","tokens_in":21493,"tokens_out":3037,"duration_ms":30760,"significance":"If the reported results are correct, DH-Mamba is a practically relevant contribution: it demonstrates that a Mamba-based architecture can be effectively adapted to the k-space domain, and the ablations support the contribution of each proposed component (circular k-space scan, hierarchical scanning, and local enhancement). The efficiency numbers are attractive, and the qualitative improvements shown are consistent with the quantitative gains on CC359 and fastMRI. However, the multi-coil results on SKM-TEA contain a physically impossible NMSE value that undermines the claim of consistent superiority until it is corrected and the surrounding numbers are re-verified. The paper does not yet release code, so the numbers cannot be independently checked.","major_comments":[{"comment":"The NMSE entry for DH-Mamba at 8x acceleration factor is internally impossible. NMSE is non-negative, so the reported mean of 0.0020 with standard deviation 0.0032 implies a negative lower tail, and the mean is roughly eight times smaller than the AF=4 NMSE (0.0156) despite a lower PSNR (32.97 dB vs 35.43 dB). This is almost certainly a transcription error (e.g., 0.0202 or 0.0220), but as printed it invalidates the reported AF=8 multi-coil comparison and weakens the claim that DH-Mamba 'consistently surpasses' prior methods on SKM-TEA. The authors must correct this entry, re-check all metrics in Table III, and reperform the comparison against ReconFormer (whose AF=8 NMSE is 0.0239) once the correct value is established.","section":"Table III (SKM-TEA)"}],"minor_comments":[{"comment":"The text says 'a circular scanning scheme is deigned'—this should read 'designed'.","section":"Section III-C1"},{"comment":"The standard deviation for DH-Mamba SSIM at AF=8 (0.0025) is an order of magnitude smaller than at AF=4 (0.0232) and also much smaller than the corresponding values for other methods; please verify that this is not a typographical error.","section":"Table III"},{"comment":"The caption states 'A lager ERF is indicated'—this should read 'A larger ERF'.","section":"Figure 9 caption"},{"comment":"The name of the transformer baseline is written inconsistently as 'ReconFormer' and 'Reconformer'; please standardize the spelling.","section":"Throughout"},{"comment":"The paper claims to 'pioneer vision Mamba in k-space learning' while citing MambaMIR and MMR-Mamba; please clarify explicitly how the proposed k-space circular scanning differs from any k-space processing in those methods, since the novelty claim is currently only implicit.","section":"Related Work and Section III-C1"},{"comment":"The ablation table uses check marks to denote enabled components, but the rows do not clearly indicate which component is removed; for example, row (a) is described as removing the image branch, yet all columns appear checked. Please use explicit labels such as 'w/o Img.' to make the ablation settings unambiguous.","section":"Table V"}],"recommendation":"major_revision","confidential_remarks":"The impossible NMSE in Table III is likely a mere typo, but because code is not yet available and the multi-coil claim rests on this table, the authors must correct the value and re-evaluate the comparison. If the correct value is around 0.020, the AF=8 advantage over ReconFormer shrinks from an order-of-magnitude gap to a modest one, so the abstract's wording about 'significant' improvement on all datasets may need softening. I would also encourage the editor to treat the fastMRI AF=4 improvement (0.14 dB PSNR) as not clearly significant without a paired statistical test."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:2501.08163. First, the architecture work is genuinely decent: dual-domain Mamba with circular k-space scanning, hierarchical HR/LR scanning, and a local enhancement module, all ablated carefully. Second, Table III contains an impossible number that currently blocks me from trusting the multi-coil conclusion.\n\nWhat is actually new: the circular scanning path tailored to k-space's concentric frequency structure, and the hierarchical scanning that runs one HR path plus three LR paths to cut sequence length and compute. The ablations in Tables V and VI show each component earns its keep, especially the k-scan (0.56 dB over vanilla SS2D) and LEM. The efficiency comparison is favorable: 117 GFLOPs and 1.87 M parameters against ReconFormer's 342 GFLOPs. The citation pattern looks fine; they acknowledge MambaMIR and MMR-Mamba, and the Discussion honestly flags Mamba's causal modeling limitation.\n\nThe soft spot is load-bearing. In Table III, DH-Mamba's NMSE at 8x AF on SKM-TEA is reported as 0.0020 ± 0.0032. That is physically impossible: a non-negative metric cannot have a mean 2.5 standard deviations below zero, and it is eight times better than their own 4x NMSE (0.0156) despite a lower PSNR. This is almost certainly a transcription error, but without released code or a reproducibility appendix, the SKM-TEA multi-coil claim rests on an impossible number. If the intended value is around 0.020, the edge over ReconFormer at 8x shrinks substantially. I also note the 'significantly outperforms' wording holds for CC359 (3+ dB over FMTNet) but is thinner on fastMRI (0.14–0.35 dB), so the claim is uneven. The hand-designed circular scan lacks a formal justification, but the ablation already shows it contributes; I consider that a minor point.\n\nThis deserves peer review, but only after the authors correct Table III and ideally release code. I'd send it to a serious referee with a note to check the SKM-TEA numbers. If the anomaly is a typo, the paper is a solid incremental contribution to Mamba-based MRI reconstruction.","headline":"Solid Mamba-for-MRI architecture paper with careful ablations, but the impossible NMSE in Table III undercuts the multi-coil claim until fixed.","tokens_in":22012,"tokens_out":1770,"would_cite":false,"duration_ms":17507,"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 dual-domain Mamba that scans k-space in concentric circles reconstructs undersampled MRI to 39.32 dB PSNR at 4x acceleration, beating prior methods by more than 3 dB at lower computational cost.","keywords":["MRI reconstruction","Mamba","state space model","k-space","circular scanning","hierarchical scanning","compressed sensing","accelerated MRI"],"falsifier":"Train the k-space branch with the circular scan replaced by a random fixed permutation of the same tokens, a spiral-from-center order, or a standard row-and-column scan while keeping every other component identical; if PSNR on CC359 at 4x stays within a few tenths of a decibel of 39.32, the claim that frequency-ordered circular unfolding is essential would be falsified. Likewise, if the gain over the best transformer baseline shrinks or vanishes on a dataset with non-Cartesian sampling, the k-space ordering claim would be pattern-specific rather than general.","tokens_in":21000,"feed_emoji":"🩻","tokens_out":6769,"duration_ms":59785,"temperature":0.7,"pith_summary":"This paper claims that adapting the Mamba sequence model to the structure of MRI data, processing both the image and its Fourier-domain k-space spectrum, makes accelerated MRI reconstruction substantially more accurate and cheaper than current CNN, transformer, and vanilla Mamba approaches. The method, DH-Mamba, scans k-space along concentric circular paths from low to high frequency rather than row by row, uses a hierarchical scan that keeps one high-resolution path and three downsampled paths to avoid long-range forgetting, and adds a pixel-wise gating module for local feature diversity. On the CC359 brain dataset at 4x acceleration it reports 39.32 dB PSNR, more than 3 dB above the best prior method, with 117 GFLOPs and 1.87M parameters, and it also leads on fastMRI and SKM-TEA. If correct, the work suggests that the ordering of tokens fed to a state space model should respect the native structure of the signal, not just the image grid.","feed_headline":"Circular k-space scan lifts MRI reconstruction by 3 dB","feed_subtitle":"Dual-domain Mamba hits 39.32 dB at 4x acceleration on CC359 while cutting FLOPs to 117G.","key_machinery":"The load-bearing object is DH-Mamba, a dual-branch network whose design is organized around how MRI data is arranged. In the k-space branch, the key mechanism is circular scanning: instead of unfolding a 2D spectrum row by row, the spectrum is traced along four concentric, frequency-ordered paths so that low frequencies at the center are read before high frequencies at the periphery, preserving the concentric structure of k-space and giving the S6 state space model a sequence whose order encodes spectral proximity. In both branches, hierarchical scanning divides the four scan directions into one high-resolution path and three low-resolution paths with $s=2$ downsampling, processed by separate S6 blocks and upsampled back, which shortens sequence length to counter long-range forgetting and reduces cost. The local enhancement module replaces the usual MLP with a pixel-wise gating mask multiplied against local convolution features, reintroducing spatially varying detail. Together these mechanisms let the model keep a global receptive field at linear complexity.","core_discovery":"On its own terms, the central discovery is that Mamba's selective state space mechanism works for MRI reconstruction when it is applied in both the image domain and the k-space domain, with scans designed for each domain. The k-space branch Fourier-transforms features, unfolds the spectrum along four concentric circular paths ordered from low to high frequency, processes those sequences with S6 blocks, and transforms back; the image branch uses standard row and column scans. A hierarchical scan splits the four directions so only one operates at full resolution while the other three run on downsampled maps, shortening sequences and reducing the forgetting of early tokens. A local enhancement module multiplies convolution-derived features by a learnable pixel-wise gate to restore spatial variation that Mamba's linear aggregation suppresses. The paper reports consistent gains over previous state of the art across three public datasets, multiple acceleration factors, and Cartesian, radial, and random undersampling masks, with lower FLOPs and parameters than the strongest baselines.","pith_inferences":["Editorial inference: circular scanning is a general recipe, not an MRI-specific trick; any inverse problem whose signal lives in a Fourier-like or polar-ordered domain could benefit from ordering tokens by frequency radius instead of raster order.","Editorial inference: the paper's own discussion admits that Mamba's causality is a ceiling, so a bidirectional or non-causal variant that can read both earlier and later tokens along the circular path is a natural next step and might close the remaining gap to ground truth.","Editorial inference: the 3:1 LR/HR path ratio is chosen empirically, and the optimal ratio likely depends on image resolution and anatomy, so a learned or adaptive allocation of scan paths across scales is a testable extension.","Editorial inference: the claim to pioneer Mamba in k-space should be read narrowly; the novelty is the circular scan and the dual-domain combination, not the use of Mamba itself, and the most informative future comparison is against other k-space-aware SSM scans under identical training budgets."],"forward_implications":["If correct, DH-Mamba establishes that Mamba-based architectures can beat transformer-based ones on accelerated MRI, not just match them, while using less computation than both.","The ablation showing a 0.98 dB drop when the k-space branch is removed implies that explicitly modeling the frequency layout of k-space is itself worth close to a decibel of PSNR.","The chosen 3:1 ratio of low-resolution to high-resolution scan paths is presented as the best trade-off point: four high-resolution paths cost 203 GFLOPs and give 39.14 dB, while four low-resolution paths drop to 38.27 dB, so the hierarchy balances detail against long-range forgetting.","The method reports gains across Cartesian, radial, and random masks and across single-coil and multi-coil data, which would make the design independent of any one sampling geometry."],"supporting_citations":[{"why":"Supplies the S6 selective state space mechanism that every branch of the model runs on.","marker":"[21]"},{"why":"The Mamba image-restoration baseline whose architecture and SS2D scanning the paper adapts and compares against.","marker":"[25]"},{"why":"The recurrent-transformer baseline that defines the SKM-TEA train/validation/test split and is the strongest competing method on that dataset.","marker":"[13]"},{"why":"Provides the fastMRI dataset and the Cartesian undersampling mask generation used in the main comparisons.","marker":"[60]"},{"why":"Cited as the source of the CC359 brain dataset used for the headline 39.32 dB result.","marker":"[62]"},{"why":"Provides the multi-coil SKM-TEA dataset used to test generalization beyond single-coil data.","marker":"[63]"},{"why":"Documents long-range forgetting in Mamba, the problem the hierarchical scan is designed to solve.","marker":"[27]"},{"why":"Introduces hierarchical high/low-resolution scanning in vision Mamba that the paper's Hi-scan builds on.","marker":"[28]"},{"why":"A powerful Mamba restoration baseline whose poor MRI results motivate the k-space-specific design.","marker":"[54]"}],"fun_headline_variants":["Dual-domain Mamba with circular k-space scan speeds MRI reconstruction","Hierarchical Mamba cuts MRI recon FLOPs while improving PSNR","Circular scanning in k-space boosts Mamba for MRI reconstruction","Mamba MRI recon: dual-domain hierarchy beats state-of-art at lower cost","K-space Mamba: circular scans and hierarchy advance MRI reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that unfolding k-space along concentric frequency-ordered circular paths is a better input order for Mamba than row-and-column scans, but the paper gives no proof that this ordering is optimal, only an ablation showing it helps by 0.56 dB; if the benefit came instead from the extra k-space branch or the hierarchical downsampling, the circular-scan argument would not carry the result.","fun_headline_variants_meta":{"raw":{"variants":["Dual-domain Mamba with circular k-space scan speeds MRI reconstruction","Hierarchical Mamba cuts MRI recon FLOPs while improving PSNR","Circular scanning in k-space boosts Mamba for MRI reconstruction","Mamba MRI recon: dual-domain hierarchy beats state-of-art at lower cost","K-space Mamba: circular scans and hierarchy advance MRI reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2959,"prompt_tokens":1052,"completion_tokens":1907,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":1814}},"tokens_in":668,"tokens_out":1907,"duration_ms":13153,"temperature":1.0,"reasoning_tokens":1814,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:30:10.112809+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the k-space branch with the circular scan replaced by a random fixed permutation of the same tokens, a spiral-from-center order, or a standard row-and-column scan while keeping every other component identical; if PSNR on CC359 at 4x stays within a few tenths of a decibel of 39.32, the claim that frequency-ordered circular unfolding is essential would be falsified. Likewise, if the gain over the best transformer baseline shrinks or vanishes on a dataset with non-Cartesian sampling, the k-space ordering claim would be pattern-specific rather than general.","supporting_citations":[{"cited_title":"Reconformer: Ac- celerated mri reconstruction using recurrent transformer,","cited_arxiv_id":null,"evidence_quote":"The recurrent-transformer baseline that defines the SKM-TEA train/validation/test split and is the strongest competing method on that dataset."},{"cited_title":"Efficient High-Resolution Visual Representation Learning with State Space Model for Human Pose Estimation","cited_arxiv_id":"2410.03174","evidence_quote":"Introduces hierarchical high/low-resolution scanning in vision Mamba that the paper's Hi-scan builds on."},{"cited_title":"Vmambair: Visual state space model for image restoration,","cited_arxiv_id":null,"evidence_quote":"A powerful Mamba restoration baseline whose poor MRI results motivate the k-space-specific design."}],"review_version":1}