{"id":"acd1d56f-99df-4e8d-96a4-b8c915bdc4f6","arxiv_id":"2504.18520","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"RSFR combines a Mamba-based coarse reconstruction, zero-shot SAM myocardial priors, and semantic feature fusion to improve undersampled cardiac diffusion MRI reconstruction and downstream DTI accuracy.","lead":"RSFR is a three-stage deep-learning pipeline that reconstructs undersampled cardiac diffusion MRI: it first makes a coarse image, uses the Segment Anything Model to locate the heart muscle, then fuses that anatomical mask into a refinement network to sharpen the final image. On 457 retrospective cardiac DTI cases it reports higher image fidelity and lower errors in diffusion-tensor parameters than six comparison methods at acceleration factors 2, 4, and 8.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SAM mask quality is never measured even though manual segmentations are available; if the zero-shot masks mislocate myocardium, the semantic-refinement mechanism, which is RSFR's core novelty, remains unverified.","rationale":"The paper's central claim is that RSFR achieves state-of-the-art reconstruction quality and accurate DT parameter estimation under high undersampling. The empirical numbers in Table 1 and Fig. 4 are internally consistent and support the claim as reported. The weakest link is the mechanism that makes RSFR distinctive: the coarse-to-fine use of SAM semantic priors. Equation (2) and §3.3 assume that the top-scoring SAM masks generated from the coarse reconstruction are reliable myocardial priors, and the SFI module is designed to let those priors guide refinement. The paper never measures mask accuracy, even though manual segmentations are available and are used in the Fig. 7 ablation. That ablation shows ideal manual masks outperform SAM masks, which is consistent with the SAM masks being suboptimal. Without a direct mask-quality measurement, one cannot distinguish between genuine semantic guidance and the effect of adding a second refinement network with more parameters. This is a load-bearing concern because it targets the claimed novelty, not the raw performance numbers. The proposed check is concrete and inexpensive: compute overlap metrics between SAM masks and the already-available manual masks. If the overlap is high, the concern lands only weakly; if it is low, the paper's explanation of why RSFR works should be revised. The reader's conditional verdict already accounts for this missing evidence, so no verdict change is needed.","tokens_in":16425,"tokens_out":4149,"duration_ms":46769,"concrete_test":"Using the manual myocardium masks already employed in the w/ Ref. Mask ablation (Fig. 7), compute Dice/IoU and centroid distance between the top SAM masks from §3.3 and the manual masks for the 23 test cases at AF×2, ×4, and ×8. If average Dice is high (>0.8) and the w/ Ref. Mask advantage over w/ SAM is small, the concern is resolved; if Dice is low or the SSIM/FA gap is large, the semantic-prior claim needs weakening.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the coarse Mamba reconstruction is reliable enough for SAM to produce useful myocardial semantic priors, and that those priors are what drive the refinement gains. In Eq. (2), Fseg = HS(¯x), and §3.3 states that the top three predictions with the highest scores are used; the SFI module then concatenates these masks with reconstruction features and applies channel-wise attention. Nowhere in the paper is the accuracy of these masks checked against ground-truth myocardium, although manual segmentations are available, since they are used in the w/ Ref. Mask arm of Fig. 7. The ablation shows w/ Ref. Mask consistently gives the best SSIM at every AF and best FA MAE at AF×8, which is exactly what one would expect if the SAM-derived priors are imperfect. If the masks frequently select non-myocardial regions, the SFI attention can amplify the wrong structures, and the reported image and DT gains could be attributed to the extra refinement subnetwork rather than to semantic guidance. This does not contradict any internal result, but it leaves the central mechanism unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RSFR, a coarse-to-fine reconstruction framework for cardiac diffusion-weighted MRI that combines a Mamba-based coarse reconstruction backbone, zero-shot semantic priors from the Segment Anything Model (SAM), and a fusion-and-refinement subnetwork with a Semantic Feature Integration module. The method is evaluated on 457 retrospectively acquired cardiac DTI cases at acceleration factors ×2, ×4, and ×8, with comparisons against six baseline/SOTA methods using DWI quality metrics (SSIM, PSNR, LPIPS) and downstream diffusion-tensor parameters (MD, FA, HA). Ablations address the segmentation model choice and the reconstruction backbone. The central claim is that RSFR achieves state-of-the-art reconstruction fidelity and the most accurate DT parameter estimation under high undersampling.","tokens_in":16596,"tokens_out":5238,"duration_ms":50089,"significance":"If the empirical claims are correct, RSFR offers a practical semantic-aware pipeline for cardiac DTI reconstruction that improves both image fidelity and tensor-parameter accuracy while requiring no manual annotations at inference. The evaluation is substantial: 457 cases spanning multiple pathologies, three undersampling factors, six comparators, per-case standard deviations, and Mann-Whitney significance tests for the image-domain metrics. The ablation design, including the w/ Ref. Mask arm, is informative. However, the strength of the DT-parameter claim is currently limited by the absence of significance testing, and the central semantic-prior mechanism lacks direct validation against ground-truth myocardium masks.","major_comments":[{"comment":"The central claim that RSFR achieves the lowest MAE for MD, FA, and HA is not supported by any statistical test or measure of variability; the text reports comparisons of point estimates only, and no error bars, confidence intervals, or p-values are provided for the DT metrics. Because quantitative accuracy is the primary clinical motivation and the test set has only 23 cases, this is a load-bearing omission. The authors should report per-case distributions and pairwise significance tests (e.g., Mann-Whitney or Wilcoxon) for the DT metrics, analogous to the stars in Table 1.","section":"§4.4.2, Fig. 4"},{"comment":"The semantic-prior mechanism is not validated. The paper never measures the accuracy of SAM's zero-shot masks against the manual myocardium segmentations that are clearly available, since those manual masks are used in the w/ Ref. Mask ablation of Fig. 7. The ablation shows that w/ Ref. Mask achieves the best SSIM at all AFs and the best FA MAE at AF×8, which is exactly the pattern expected if SAM masks are suboptimal. Without reporting mask overlap (e.g., Dice) and analyzing the effect of mask errors on the refined reconstruction, the claim that semantic guidance drives the improvement remains unverified; the gains could come from the additional refinement subnetwork alone.","section":"§3.3, Fig. 7"},{"comment":"The comparison against SOTA methods is incomplete because the paper does not state whether the baseline methods (D5C5, DAGAN, STGAN, UNet, SwinMR, MambaMIR) were retrained on the same 434-case training set with the same undersampling simulation, or whether pre-trained weights were used. This is essential for a fair comparison; if baselines were not trained on the same data distribution, the reported performance differences could reflect training conditions rather than algorithmic superiority. The implementation details section only describes the RSFR training protocol and omits any information about baseline training or checkpoint sources.","section":"§4.4, §4.2"},{"comment":"The SAM inference procedure is not described with sufficient specificity for reproducibility. SAM typically requires input prompts and is designed for RGB images, but the paper does not state how prompts were generated, whether the DWI images were converted to three channels, how the 'top three predictions with the highest scores' were selected, or what thresholds or post-processing were applied to the masks. Since SAM is a core component of the proposed mechanism, these details should be provided.","section":"§3.3, §4.2"}],"minor_comments":[{"comment":"Two references are incomplete: 'MambaIRv2 ?' and 'CU-Mamba ?' appear in the related-work section without citation keys or bibliographic entries.","section":"§2.2"},{"comment":"The notation for the coarse reconstruction is inconsistent: Eq. (2) uses \\bar{x}, while §3.3 refers to it as \\hat{x}_coarse; Eq. (1) also uses \\hat{x} for the generic inverse-problem solution, which later denotes the final refined output.","section":"§3.1, §3.3"},{"comment":"The final sentence of the conclusion contains a redundant phrase: 'RSFR has represented a significant advancement in cDTI reconstruction, positioning RSFR as a significant advancement in cDTI reconstruction' should be simplified.","section":"§6"},{"comment":"The paper would benefit from a data and code availability statement, particularly because the post-processing pipeline is described as relying on the open-source INDI tool but no link is provided for the RSFR implementation itself.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript cites a large number of works by the same authors (e.g., Huang et al. 2022a, 2022b, 2024a, 2024b, 2025; Wang et al. 2024a, 2024b). This is not improper per se, but the authors should ensure that the novelty of RSFR is clearly positioned relative to their own prior Mamba-based reconstruction method (MambaMIR, Huang et al. 2025), since the current backbone is described as inspired by it. The editor may also wish to verify the intended submission status, as the article header indicates 'Medical Image Analysis (2025)' while the arXiv version is v1 dated April 2025."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RSFR is a genuinely useful empirical contribution: a coarse-to-fine cDTI reconstruction pipeline that runs SAM on the coarse Mamba reconstruction rather than on the degraded input, fuses the semantic priors through a channel-attention module, and trains end-to-end with downstream tensor-parameter evaluation. The components are known, but the assembly is new and sensible, and the paper gives it a fair test: six baselines, 457 cases, per-case standard deviations, Mann-Whitney tests on the DWI metrics, and ablations on both the segmentation model and the backbone. The DT-parameter evaluation (MD, FA, HA) is the right thing to care about for this application, and the reported gains are consistent with the image-fidelity numbers. I believe the central empirical claim—that RSFR improves reconstruction fidelity and DT accuracy at AF2/4/8—is defensible on the evidence shown.\n\nThe soft spots are real but mostly addressable. The loss coefficients alpha, beta, gamma in Eq. (3) are never given, and no code or weights are released, so the results are not independently reproducible. The DT-parameter MAE results in Fig. 4 have no significance tests or error bars; the claims of 'lowest MAE' rest on point estimates. The test set is only 23 cases, and the clinical-translation language outruns that. The most substantive issue is the one the stress-test flags: the paper never measures SAM mask accuracy against the manual segmentations it clearly has (they are used for the w/ Ref. Mask arm of Fig. 7). That ablation shows ideal masks beat SAM masks on SSIM at every AF and on FA at AF8, which is consistent with SAM being imperfect. The w/ SAM vs N.A. comparison does show that adding the semantic branch helps, so the mechanism isn't vacuous, but 'semantic-aware' specifically is not directly validated. This is a gap in the paper's central narrative, not a contradiction, and it can be closed with a simple mask-quality experiment.\n\nWho is this for? Anyone working on MRI reconstruction, especially cardiac DTI or semantic-prior-guided restoration, will get value from it. It deserves a serious referee. For peer review, I'd recommend major revision: release code and specify loss weights, add significance testing or at least error bars for the DT parameters, and report SAM mask Dice/Jaccard against the manual masks. The way it stands, it is a good methods paper with a poorly verified mechanism.","headline":"A solid, well-engineered cDTI reconstruction paper whose central semantic-prior mechanism is never directly validated because SAM mask quality is never measured.","tokens_in":17264,"tokens_out":3229,"would_cite":true,"duration_ms":32225,"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 deriving semantic priors from a coarse Mamba reconstruction rather than from undersampled input lets a refinement network improve both cardiac DWI fidelity and diffusion tensor accuracy.","keywords":["cardiac diffusion tensor imaging","MRI reconstruction","undersampled MRI","Segment Anything Model","Mamba","semantic-aware reconstruction","coarse-to-fine","diffusion tensor parameters"],"falsifier":"Measure Dice overlap between SAM's top-three masks and manual myocardium segmentations on the coarse reconstructions, then rerun refinement with deliberately misplaced masks; if overlap is low or misplaced masks do not degrade SSIM and DT MAE, the semantic prior is not what carries the improvement.","tokens_in":16156,"feed_emoji":"🫀","tokens_out":8938,"duration_ms":80742,"temperature":0.7,"pith_summary":"The paper proposes RSFR, a coarse-to-fine pipeline for cardiac diffusion-weighted MRI that first reconstructs a coarse image with a Vision Mamba network, derives myocardial semantic masks from that coarse image using the Segment Anything Model, and then refines the image while fusing the masks through channel attention. Its central claim is that taking the semantic prior from the coarse reconstruction, rather than from the severely undersampled input, makes the prior reliable enough to improve both pixel-level fidelity and downstream diffusion-tensor measurements. Across acceleration factors 2, 4, and 8, RSFR reports higher SSIM and PSNR than the compared methods and the lowest mean absolute errors for mean diffusivity, fractional anisotropy, and helix angle in nearly all settings. If correct, this would make high-acceleration cardiac DTI more clinically practical, where repeated breath holds and low signal-to-noise ratio currently limit adoption.","feed_headline":"RSFR beats prior methods in cardiac diffusion MRI at 8x undersampling","feed_subtitle":"Combining a Mamba backbone with zero-shot SAM masks improves image fidelity and tensor measurements.","key_machinery":"The load-bearing machinery is the coarse-to-fine equation $\\bar{x}=H_R(A^H y)$, $F_{seg}=H_S(\\bar{x})$, $\\hat{x}=H_{FR}(\\bar{x},F_{seg})$, in which the semantic prior is deliberately obtained from the coarse reconstruction rather than from the undersampled input. $H_R$ and $H_{FR}$ are U-shaped Vision Mamba networks built from Visual State Space blocks that use a cross-scan mechanism to flatten image patches into four directed sequences, giving global context with linear complexity. $H_S$ is the frozen Segment Anything Model with ViT-H weights, whose top three highest-scoring masks are used without training or manual annotation. The Semantic Feature Integration module concatenates reconstruction and mask features, aligns them with convolution, instance normalization, and GeLU, and applies channel-wise attention so the refinement can selectively enhance myocardial regions and suppress background.","core_discovery":"The paper's discovery, stated on its own terms, is that semantic-aware coarse-to-fine reconstruction improves cardiac DWI fidelity and quantitative tensor accuracy. The pipeline is written as $\\bar{x}=H_R(A^H y)$, $F_{seg}=H_S(\\bar{x})$, $\\hat{x}=H_{FR}(\\bar{x},F_{seg})$: a Mamba-based reconstruction model produces a coarse image, a frozen SAM produces top-three masks from that coarse image, and a second Mamba-based model fuses reconstruction and mask features through Semantic Feature Integration modules. On a 457-case dataset spanning healthy hearts and six disease groups, RSFR achieves the best SSIM and PSNR at all tested acceleration factors and the lowest MAE for global mean MD, FA, and HA gradient in nearly all comparisons, with the largest visible gap at acceleration factor 8. The authors take this as evidence that semantic priors extracted from a coarse reconstruction suppress non-myocardial contributions while preserving the subtle intensity differences that tensor estimation depends on.","pith_inferences":["The ablation gap between SAM masks and reference masks at AF8 implies RSFR's current semantic prior is not the ceiling; a mask that localizes the myocardium more precisely could improve high-acceleration DTI further.","The paper's observed link between PSNR/SSIM and DT MAE, alongside a weak link with LPIPS, suggests that for quantitative MRI, perceptual metrics may reward clinically misleading reconstructions; fidelity-based selection is safer for tensor estimation.","The same coarse-to-fine semantic prior design should transfer to other quantitative MRI settings, such as T1/T2 mapping or brain diffusion, where ROI-specific intensity accuracy matters, but this would need separate validation because SAM's masks may be less reliable in more complex anatomy.","A direct test of the mechanism would be to compare the top-three SAM masks with the single highest-confidence mask or confidence-weighted masks; improved results would show mask selection, not fusion, is the current bottleneck."],"forward_implications":["At acceleration factor 8, RSFR reports SSIM 0.754 and PSNR 24.25, the best among compared methods, with sharper myocardial walls and lower residual errors.","Global mean MD, FA, and HA gradient show lower MAE under RSFR than the compared methods in nearly all settings, with narrower interquartile ranges for FA.","Removing the segmentation stage worsens SSIM at every acceleration factor, so the semantic prior is a necessary part of the reported gain.","Because SAM requires no task-specific training or annotations, the framework can be applied to new scanners or protocols without building a segmentation dataset."],"supporting_citations":[{"why":"Supplies the frozen Segment Anything Model that generates the zero-shot semantic masks from the coarse reconstruction.","marker":"Kirillov et al. (2023)"},{"why":"VMamba provides the Visual State Space backbone and cross-scan mechanism used by the reconstruction and refinement models.","marker":"Liu et al. (2024b)"},{"why":"MambaMIR is the Mamba-based medical image reconstruction baseline and an architectural inspiration for the backbone.","marker":"Huang et al. (2025)"},{"why":"Demonstrates that semantic priors distilled from coarsely recovered images improve restoration, motivating extraction of SAM priors from the coarse reconstruction.","marker":"Zhang et al. (2024)"},{"why":"Establishes the cDTI acquisition and post-processing protocol used to estimate diffusion tensors and compute MD, FA, and HA.","marker":"Ferreira et al. (2014)"},{"why":"fastMRI provides the Cartesian undersampling mask protocol used to simulate acceleration factors 2, 4, and 8.","marker":"Zbontar et al. (2018)"},{"why":"Earlier comparison study of deep learning cDTI reconstruction that frames the task and confirms the need for pixel-faithful reconstruction.","marker":"Huang et al. (2024a)"}],"fun_headline_variants":["Coarse-to-fine with SAM masks refines cardiac diffusion MRI","RSFR: semantic priors sharpen cardiac tensor estimates","Mamba reconstruction plus SAM masks: cardiac DTI gains","Better cardiac DTI from coarse-to-fine semantic refinement","Zero-shot SAM lifts cardiac diffusion MRI fidelity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"RSFR's reported gains rest on the assumption that SAM's zero-shot masks, taken from the coarse reconstruction, actually locate the myocardium; the paper never measures mask accuracy against manual segmentations, and its own ablation shows hand-drawn reference masks outperform SAM masks at acceleration factor 8.","fun_headline_variants_meta":{"raw":{"variants":["Coarse-to-fine with SAM masks refines cardiac diffusion MRI","RSFR: semantic priors sharpen cardiac tensor estimates","Mamba reconstruction plus SAM masks: cardiac DTI gains","Better cardiac DTI from coarse-to-fine semantic refinement","Zero-shot SAM lifts cardiac diffusion MRI fidelity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3444,"prompt_tokens":912,"completion_tokens":2532,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":2453}},"tokens_in":528,"tokens_out":2532,"duration_ms":17821,"temperature":1.0,"reasoning_tokens":2453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:14:28.260747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure Dice overlap between SAM's top-three masks and manual myocardium segmentations on the coarse reconstructions, then rerun refinement with deliberately misplaced masks; if overlap is low or misplaced masks do not degrade SSIM and DT MAE, the semantic prior is not what carries the improvement.","supporting_citations":[],"review_version":1}