{"id":"1f222d6c-e01f-44c8-8fca-57b69235482a","arxiv_id":"2412.10776","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"FPS-Former, a Vision Transformer with frequency modulation, spatially purified attention, and scale-diversified feed-forward blocks, outperforms prior MRI reconstruction methods on CC359, fastMRI, and SKM-TEA.","lead":"FPS-Former is a Vision Transformer for MRI reconstruction that adds frequency-aware attention, grouped token attention, and multi-scale feed-forward paths. The model reports better image quality than previous methods on three public MRI datasets while using similar or fewer computations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 contradicts the paper's own claim: on fastMRI AF=4, ReconFormer's PSNR (32.53) is higher than FPS-Former's (32.51), so 'consistently surpasses' and 'outperforms' are unsupported as written.","rationale":"I focused on the internal consistency of the reported evidence rather than on the speculative possibility of unfair baselines. The reader's weakest assumption (baseline protocol) is important and related, but the paper's own Table 1 already contains a concrete counterexample to the unqualified claim: on fastMRI AF=4, ReconFormer's PSNR is 0.02 dB higher than FPS-Former's. Since PSNR is one of the three reported metrics, 'consistently surpasses' and the abstract's 'outperforms' are not supported by the paper's own data. This is the single most load-bearing concern because the entire contribution rests on empirical superiority. The proposed test, re-running with identical protocol and repeated seeds, simultaneously addresses both the internal contradiction and the reader's protocol concern. I do not think this requires changing the conditional verdict, but it makes the revision requirement sharper: the authors must either provide statistical significance showing the difference is noise or qualify the claim. Agreement with the reader is partial: they pointed at external verification; I point at an internal contradiction in the tables.","tokens_in":13719,"tokens_out":12237,"duration_ms":114433,"concrete_test":"Independently re-run the fastMRI AF=4 comparison with identical Cartesian masks, data splits, loss, and training schedule for both FPS-Former and ReconFormer across at least three seeds, and report mean±std for NMSE, SSIM, and PSNR. If ReconFormer's mean PSNR is not below FPS-Former's, the abstract and Section 5 must be revised to qualify the superiority claim by metric and uncertainty, and the 'consistently surpasses' sentence must be removed or made precise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: FPS-Former 'outperforms state-of-the-art methods while requiring lower computational costs.' The most direct support is Tables 1-2. In Table 1, fastMRI, AF=4, ReconFormer has NMSE 0.0320, SSIM 0.7327, PSNR 32.53; FPS-Former has NMSE 0.0316, SSIM 0.7337, PSNR 32.51. Thus on a primary reported metric, PSNR, the leading transformer baseline beats FPS-Former. The text in Section 5 states FPS-Former 'consistently surpasses other Transformer-based approaches across different acceleration rates on both datasets,' and the abstract makes an unqualified 'outperforms' claim. The paper's own table therefore fails to support the claim as stated. No error bars or repeated runs are reported, so it is impossible to tell whether the 0.02 dB deficit is noise; but an unsupported universal claim in the abstract and text is a real correctness and overclaiming issue, not merely a missing experiment. This is the most load-bearing concern because it sits in the evidence for the headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FPS-Former, a ViT-based MRI reconstruction framework designed to address three issues: high-frequency information loss in self-attention, attention over content-irrelevant tokens, and lack of multi-scale feature modeling in feed-forward networks. It introduces three modules—Frequency Modulation Attention Module (FMAM), Spatial Purification Attention Module (SPAM), and Scale Diversification Feed-forward Network (SDFN)—plus a Hybrid Experts Feature Refinement (HEFR) module. The method is evaluated on CC359, fastMRI, and SKM-TEA under Cartesian, radial, and random undersampling masks at acceleration factors 4x, 5x, 8x, and 10x. The central claim is that FPS-Former outperforms state-of-the-art methods while requiring lower computational costs.","tokens_in":14051,"tokens_out":6293,"duration_ms":53112,"significance":"If the results are reproducible, the paper makes a useful architectural contribution by explicitly targeting known ViT weaknesses in low-level vision and MRI reconstruction. The ablation study in Table 4 supports the contribution of each module, and the efficiency analysis (152G FLOPs vs 342G for ReconFormer in Table 6) is a practical plus. The motivation for FMAM is grounded in prior observations about MSA as a low-pass filter, and the qualitative frequency-response analysis in Figure 4 provides supporting evidence. However, the significance is tempered by unresolved questions about protocol fairness, statistical robustness, and the accuracy of the paper's own performance claims.","major_comments":[{"comment":"The abstract's unqualified claim that FPS-Former 'outperforms state-of-the-art methods' and the text's claim that it 'consistently surpasses other Transformer-based approaches across different acceleration rates on both datasets' are contradicted by Table 1 on fastMRI at AF=4, where ReconFormer achieves PSNR 32.53 vs FPS-Former's 32.51. The paper should either qualify these claims (e.g., 'generally outperforms' or 'outperforms on most metrics and datasets') or provide statistical evidence (multiple runs with error bars) showing that the 0.02 dB deficit is not noise. This is the central empirical claim, so the overstatement needs correction.","section":"Abstract; Section 'Comparison with State-of-the-arts'; Table 1"},{"comment":"The paper does not state whether baseline methods were retrained under the identical experimental protocol (same undersampling masks, data splits, loss function, and training schedule) or whether their reported numbers were taken from the original publications. Since the paper does not release code or detailed per-baseline training configurations, the claimed improvements over SwinMR, ReconFormer, Restormer, and AST could be artifacts of protocol differences. The authors should specify the protocol for each baseline and, ideally, provide code to allow third-party verification.","section":"Experimental Settings; Comparison with State-of-the-arts"},{"comment":"The hash-based grouping in SPAM relies on random variables a and b, but the paper does not specify whether these are sampled once and fixed or re-sampled each forward pass, nor does it state the number of hash functions used. This makes the method's grouping behavior unreproducible. Additionally, this locality-sensitive hashing scheme is essentially the one used in the Reformer architecture (Kitaev et al., 2020), which should be cited. Clarify the hashing details and reference the prior work.","section":"Spatial Purification Attention Module, Eq. (7)–(8)"},{"comment":"The Gaussian kernel in Eq. (3) uses the normalization constant 1/(σ√(2π)), which is the one-dimensional Gaussian normalization, applied to a two-dimensional Gaussian. For a normalized 2D kernel, the constant should be 1/(2πσ^2). If this is a typo, it should be corrected; if the intentionally non-normalized kernel is used, the effect on the Laplacian pyramid and on the attention recalibration should be discussed, because the DC component may not be preserved across scales.","section":"Frequency Modulation Attention Module, Eq. (3)"},{"comment":"All quantitative results are reported as single numbers with no error bars, confidence intervals, or multiple runs. Given that the difference with ReconFormer on fastMRI AF=4 is only 0.02 dB in PSNR, the single-run results are insufficient to support the claim of consistent superiority. The authors should report mean ± standard deviation over at least three seeds, or provide a paired statistical test across test slices, to establish that the observed differences are not random variation.","section":"Experiments; Tables 1 and 2"}],"minor_comments":[{"comment":"The caption lists '(3) We propose' but the corresponding panel is labeled '(c)'. The numbering should be consistent.","section":"Figure 1 caption"},{"comment":"In the sentence 'the number of groups N in SHAM', 'SHAM' should be 'SPAM'.","section":"Training Details"},{"comment":"The checkmark formatting in Table 4 is garbled in the text, making it difficult to see which components are present in each ablation row. Please ensure the table renders with clear ✓/✗ symbols.","section":"Table 4"},{"comment":"The description of the 1D Cartesian undersampling mask is vague ('similar to the fastMRI challenge'). Specify the exact mask type (e.g., variable-density random, center fraction, low-frequency lines) and how masks are generated for each dataset to improve reproducibility.","section":"Experimental Settings"},{"comment":"The paper defers hyperparameter analysis to the Supplementary Materials, but the arXiv submission does not appear to include supplementary materials. Either include the analysis or summarize the key findings for the key hyperparameters (M, N, r, E, D) in the main text.","section":"Analysis of Hyper-parameters"},{"comment":"The notation 'F ⋇ P' is used for the combined FMAM and SPAM but is not defined before Eq. (2). Define this operator or replace it with explicit notation to avoid confusion.","section":"Eq. (2) and Fig. 2"},{"comment":"The FMAM attention computation sums softmax attention scores over heads and then applies the summed matrix to a single V derived from Fin. This is non-standard; clarify whether the heads share the V projection or whether the summed matrix is used as a single-head attention weight.","section":"Eq. (5)–(6)"}],"recommendation":"major_revision","confidential_remarks":"The paper's contribution is potentially solid, but the central 'outperforms' claim is factually contradicted on one metric in Table 1, and the lack of code and protocol details prevents independent verification. The missing error bars are a significant problem in a field where 0.02 dB differences are often noise. The method itself, with its three targeted modules and ablations, seems defensible and could become acceptable after the claims are qualified, the experiments are strengthened with error bars, and the training protocols for baselines are clarified. The self-citations (Yang et al. 2024a, 2024b) are not used to support the central claims, so they are not a concern. The paper's fit to the journal (eess.IV) is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nWhat you should know: this is a competent, incremental architecture paper for accelerated MRI reconstruction. The three modules are not deeply original—SPAM is essentially Reformer's LSH attention repurposed, and the frequency-modulation idea builds on the known low-pass behavior of ViTs—but the combination is new for this task, the ablations are honest, and the gains are consistent across three datasets. The main thing to flag: the abstract and text overclaim. Table 1 itself shows ReconFormer with a higher PSNR on fastMRI AF=4 (32.53 vs 32.51). That contradicts \"consistently surpasses\" and the unqualified \"outperforms\" in the abstract. The gap is 0.02 dB, almost certainly within noise, and FPS-Former wins on NMSE and SSIM in that same cell, so the overall result survives. But the claim should be softened.\n\nWhat's actually new: FMAM, which recalibrates attention scores via a Laplacian pyramid, is a workable and reasonably clear idea. SPAM is hash-based grouped attention; this paper's version uses random projections, but it's Reformer's lineage and should cite it. SDFN is a multi-scale FFN, not exotic but effective. Table 4 shows each component contributes, and Table 5 gives useful comparisons against global/window MSA and alternative FFNs. The efficiency comparison is credible and relevant: 152G FLOPs versus ReconFormer's 342G, with comparable or better quality on most cells.\n\nSoft spots in proportion: (1) the overclaim, fixable in revision; (2) the missing Reformer citation, an obvious omission; (3) no code and no error bars, which is common in this subfield but makes \"significant\" hard to verify. I would not call any of these fatal.\n\nThis paper deserves a serious referee. The architecture is coherent, the evaluation is broad (three datasets, including multi-coil), and the efficiency argument is practically useful. My recommendation: major revision, primarily to fix the overclaim and the citation, and ideally to release code.\n\nBest,\n[Name]","headline":"Solid incremental architecture paper whose own Table 1 undercuts its 'consistently surpasses' claim; worth reviewing after a fix.","tokens_in":14558,"tokens_out":3271,"would_cite":true,"duration_ms":28317,"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":"FPS-Former, a ViT-based MRI reconstruction framework, reports state-of-the-art accuracy on three public datasets at 4x and 8x acceleration by addressing frequency loss, token noise, and missing multi-scale modeling, with lower compute.","keywords":["accelerated MRI reconstruction","vision transformer","frequency modulation attention","Laplacian pyramid","spatial purification attention","multi-scale feed-forward network","single-coil and multi-coil MRI","image restoration"],"falsifier":"Reproduce the comparison on CC359 at 4x acceleration under the paper's stated protocol (random 1D Cartesian masks, official data split, L1 loss, AdamW, 300K iterations) by training every baseline from its own official implementation. The central claim collapses if FPS-Former does not at least match its reported PSNR advantage over ReconFormer (34.38 vs 34.16) and AST (34.38 vs 32.78) under those identical conditions.","tokens_in":13541,"feed_emoji":"🧠","tokens_out":11191,"duration_ms":88614,"temperature":0.7,"pith_summary":"FPS-Former is a ViT-based MRI reconstruction framework built on the claim that the remaining gap in accelerated MRI comes from three architectural habits of vision transformers, not from data or scale. Self-attention acts as a low-pass filter, dense attention mixes content-irrelevant tokens, and the standard feed-forward network cannot represent multi-scale structure. The paper answers each with a targeted module—Laplacian-pyramid frequency recalibration, hash-based grouped attention, and multi-scale depthwise-convolution feed-forward paths—and reports state-of-the-art PSNR/SSIM/NMSE on CC359, fastMRI, and SKM-TEA at 4x and 8x acceleration while using less computation than its strongest rivals. If the reported comparisons are fair, the result is a practical recipe for transformer-based reconstruction that is both sharper and cheaper.","feed_headline":"Transformer MRI model tops three benchmarks with less compute","feed_subtitle":"FPS-Former's three-pronged transformer design beats prior models at 4x and 8x acceleration on three public datasets.","key_machinery":"The load-bearing object is the FPS block, a drop-in replacement for the standard transformer block, composed of three modules. FMAM builds a Laplacian pyramid of the input feature by subtracting progressively stronger Gaussian smoothings, computes a within-frequency attention map for each pyramid level, sums the maps, and applies the summed scores to the value computed from the original input, so high-frequency bands that plain self-attention suppresses are selectively amplified. SPAM flattens tokens, assigns each a hash code via a random projection, sorts by hash, and runs multi-head self-attention only inside each hash group, so unrelated tokens never attend to each other and quadratic cost is reduced. SDFN replaces the naive feed-forward MLP with two parallel depthwise-convolution branches ($3\\times3$ and $5\\times5$) whose outputs are cross-concatenated and refined, adding multi-scale local modeling. Around these blocks, HEFR, a self-attention-switched set of CNN experts (average pooling, separable convolutions, dilated convolutions), refines features at early and final stages, and a data-consistency layer ties the reconstruction back to the measured k-space.","core_discovery":"The central claim is that a ViT for MRI reconstruction can outperform current state of the art by repairing three internal weaknesses rather than by scaling the model. The proposed FPS-Former couples frequency modulation attention (FMAM), which computes attention scores per level of a Gaussian Laplacian pyramid and re-weights the original value features with their sum, with spatial purification attention (SPAM), which hashes tokens, sorts them, and restricts multi-head attention to hash-neighbor groups, and a scale diversification feed-forward network (SDFN) that fuses $3\\times3$ and $5\\times5$ depthwise-convolution branches. On single-coil CC359 and fastMRI and multi-coil SKM-TEA at acceleration factors 4 and 8, and under Cartesian, radial, and random masks, the authors report that FPS-Former beats CNN and transformer baselines including ReconFormer and AST, with 12.51M parameters and 152G FLOPs on CC359. The ablation study attributes the gain to all four components, with SPAM contributing the largest single PSNR drop when removed.","pith_inferences":["The hash-based grouping in SPAM is effectively a locality-sensitive hash; if it transfers, the same grouping trick could cheapen self-attention in other image-restoration and high-resolution vision tasks beyond MRI.","Because the method is architectural rather than MRI-specific, a natural testable extension is replacing the data-consistency layer with the appropriate forward operator to evaluate FPS blocks on other inverse problems, such as low-dose CT or super-resolution.","FMAM's spectral behavior could be probed directly: at intermediate layers, the frequency response of its attention maps should show measurably less high-frequency attenuation than a matched plain ViT, which is a sharper test than the end-to-end PSNR comparison.","The single L1 loss leaves room for extensions with perceptual or adversarial losses, which could improve visual fidelity even if PSNR changes little."],"forward_implications":["If the reported comparisons are correct, FPS-Former sets the new state of the art for accelerated MRI reconstruction on single-coil CC359 and fastMRI and multi-coil SKM-TEA, with larger margins at 8x than at 4x acceleration.","The ablations indicate each component carries weight independently: removing FMAM, SPAM, SDFN, or HEFR lowers CC359 PSNR at 4x by 0.24, 0.55, 0.43, and 0.40 dB respectively.","The hash-grouped attention in SPAM is the main cost saver, allowing FPS-Former to exceed ReconFormer's quality while using fewer FLOPs (152G vs 342G) and fewer parameters than AST (12.51M vs 26.10M).","The frequency and grouping modules also improve robustness across undersampling patterns, with FPS-Former leading under radial and random masks at 5x and 10x acceleration on CC359."],"supporting_citations":[{"why":"Documents that ViT self-attention behaves as a low-pass filter, motivating the frequency modulation module.","marker":"Park and Kim 2022"},{"why":"Provides the Fourier-domain analysis of oversmoothing in deep ViTs and the spectral-response evaluation the paper uses to show FMAM preserves high frequencies.","marker":"Wang et al. 2022"},{"why":"ReconFormer is the strongest ViT baseline on single-coil datasets, supplying the comparison numbers that FPS-Former must beat.","marker":"Guo et al. 2024"},{"why":"AST is the leading transformer-based restoration baseline, supplying the main efficiency and quality comparison in terms of PSNR, FLOPs, and parameters.","marker":"Zhou et al. 2024"},{"why":"Restormer is the natural-image-restoration transformer baseline that the paper equips with a data-consistency layer for fair comparison.","marker":"Zamir et al. 2022"},{"why":"Supplies the fastMRI single-coil dataset and the 1D Cartesian random under-sampling protocol used to generate inputs.","marker":"Zbontar et al. 2018"},{"why":"Supplies the SKM-TEA multi-coil dataset used to test the method beyond single-coil reconstruction.","marker":"Desai et al. 2022"},{"why":"Supplies the CC359 brain dataset and its official split, the primary benchmark for ablations and efficiency analysis.","marker":"Warfield, Zou, and Wells 2004"},{"why":"Inspires the hybrid-experts feature refinement module (HEFR) that is inserted at early and final network stages.","marker":"Chen et al. 2023"}],"fun_headline_variants":["MRI transformer fixed: frequency, purity, scale boost quality","Three-pronged ViT fix makes MRI reconstruction faster, better","FPS-Former: three attention fixes boost MRI reconstruction","ViT MRI: frequency, spatial, scale fixes yield SOTA results","FPS-Former: MRI transformer beats prior models with less compute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire comparison rests on the assumption that every baseline was trained and tested under the exact same protocol as FPS-Former; because the paper does not release code, a third party cannot independently verify that the reported baseline numbers came from that fair protocol.","fun_headline_variants_meta":{"raw":{"variants":["MRI transformer fixed: frequency, purity, scale boost quality","Three-pronged ViT fix makes MRI reconstruction faster, better","FPS-Former: three attention fixes boost MRI reconstruction","ViT MRI: frequency, spatial, scale fixes yield SOTA results","FPS-Former: MRI transformer beats prior models with less compute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"total_tokens":4229,"prompt_tokens":1045,"completion_tokens":3184,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":3096}},"tokens_in":661,"tokens_out":3184,"duration_ms":18451,"temperature":1.0,"reasoning_tokens":3096,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:36:33.340204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the comparison on CC359 at 4x acceleration under the paper's stated protocol (random 1D Cartesian masks, official data split, L1 loss, AdamW, 300K iterations) by training every baseline from its own official implementation. The central claim collapses if FPS-Former does not at least match its reported PSNR advantage over ReconFormer (34.38 vs 34.16) and AST (34.38 vs 32.78) under those identical conditions.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ReconFormer is the strongest ViT baseline on single-coil datasets, supplying the comparison numbers that FPS-Former must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AST is the leading transformer-based restoration baseline, supplying the main efficiency and quality comparison in terms of PSNR, FLOPs, and parameters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the hybrid-experts feature refinement module (HEFR) that is inserted at early and final network stages."}],"review_version":1}