REVIEW 2 major objections 5 minor
Cardiac MRI Through-Plane Super-Resolution Guided by Reference and Memory
T0 review · 2 major / 5 minor · reviewed 2026-07-09 · glm-5.2
Pith's one-line read Memory and reference views sharpen cardiac MRI through-plane
desk verdict Solid technical contribution with a real limitation in evaluation scope — synthetic degradation on a single dataset read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Coarse-to-Fine Contextual Matching (CFCM), Patch-wise Dynamic Feature Aggregation (PDFA), and Memory-based SR Propagation
What would settle it
Apply STRMSR to cardiac MRI data with realistic residual inter-slice misalignment (without prior registration correction) and measure whether the coarse-to-fine contextual matching produces correct correspondences or amplifies artifacts. If the matching degrades under misalignment, the assumption of pre-aligned input is load-bearing rather than incidental.
Extended reading notes
Core claim
The paper demonstrates that combining multi-scale correspondence refinement, patch-level adaptive fusion, and inter-slice memory propagation yields measurable improvements in through-plane cardiac MRI super-resolution, particularly at 8x upsampling where existing methods degrade most. The ablation study shows that replacing the coarse-to-fine matching with a single-scale matching strategy causes the largest performance drop (-0.60 dB PSNR), followed by removing the dynamic aggregation (-0.24 dB) and the memory bank (-0.12 dB), indicating that all three components contribute complementarily but correspondence refinement is the most critical. The performance advantage over baselines grows as 8
Load-bearing premise
The method assumes that the input low-resolution image stacks are already spatially aligned. If residual breath-hold or cardiac motion misalignment persists in clinical data, the coarse-to-fine matching could establish incorrect correspondences and degrade the output.
Editorial extensions
If this is right
- If the memory-propagation approach generalizes, the same mechanism could be applied to other anisotropic medical imaging modalities where slice-to-slice consistency is clinically important, such as fetal brain MRI or abdominal imaging.
- The finding that gains are largest at 8x upsampling—where clinical breath-hold constraints are most binding—suggests the method is most useful in exactly the regime that dominates real clinical cardiac MRI acquisition.
- The patch-wise dynamic aggregation strategy, interpreted as a content-adaptive mixture-of-experts over warped reference features, could be adopted more broadly in any multi-reference image fusion task where some references are more informative than others for a given spatial region.
- The memory bank's expansion from 2-3 frames during training to 10 frames during inference for the long-axis protocol suggests the method benefits from longer temporal context at test time, raising the question of whether even larger memory banks would yield further gains.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper proposes STRMSR, a reference- and memory-guided through-plane super-resolution framework for cardiac MRI. The method introduces three components: (1) coarse-to-fine contextual matching (CFCM) for establishing cross-view correspondence, (2) patch-wise dynamic feature aggregation (PDFA) for fusing multiple warped reference/memory features, and (3) a FIFO memory bank that propagates intermediate SR results along the through-plane axis to enforce slice-to-slice consistency. The method is evaluated on the WHS dataset under two protocols (WHS-Ortho and WHS-LAX) at ×4 and ×8 upsampling, with an ablation study isolating each component's contribution.
Significance. The problem of through-plane SR for cardiac MRI is clinically relevant, and the memory-based propagation mechanism is a reasonable adaptation of video object segmentation ideas to volumetric SR. The ablation study (Table 2) properly isolates component contributions, and the use of two reference protocols with different coverage densities provides useful diversity within the single dataset. The ×8 results on WHS-Ortho show a meaningful improvement (+0.97 dB PSNR over McMRSR), which is the most practically relevant regime. Statistical significance is reported via Wilcoxon signed-rank tests.
major comments (2)
- Section 1, Contribution (1): The paper claims that CFCM enables 'robust detail transfer between target and reference/memory images under cross-view geometric misalignment.' However, no experiment introduces residual misalignment to test this claim. In both WHS-Ortho and WHS-LAX protocols (Section 3.1), the reference views are resampled from the same underlying 3D volume with known geometric transformations and no inter-acquisition motion. The only 'misalignment' present is the exact geometric relationship between orthogonal views, which is trivially known. This is load-bearing because robustness to misalignment is stated as a primary contribution and motivates the coarse-to-fine design. The authors should either (a) add an experiment with controlled synthetic misalignment (e.g., random rigid perturbations applied to reference views) to demonstrate that CFCM degrades gracefully, or (b) re
- Table 1, WHS-LAX rows: The improvements on the more clinically realistic protocol are modest — +0.17 dB PSNR at ×4 and +0.19 dB at ×8 over McMRSR. While statistical significance is reported for ×8, the practical significance of sub-0.2 dB gains is questionable, especially given that the evaluation uses synthetic degradation on a single dataset where references are derived from the same volume. The paper should discuss this limitation explicitly and acknowledge that the stronger WHS-Ortho gains may benefit from the dense reference coverage that is less representative of clinical LAX acquisitions. Without this qualification, the abstract's claim of 'consistent improvements' overstates the WHS-LAX evidence.
minor comments (5)
- Section 3.2: The memory bank size differs between training and inference (T=2 vs T=10 for WHS-LAX). No justification or ablation is provided for this discrepancy or for the specific values chosen.
- Section 2.5 and Figure 3: The term 'temporal profile' is used for the through-plane spatial axis. While the video analogy is explained, this terminology may confuse readers expecting temporal (cardiac phase) information. Consider clarifying.
- Table 1: STRMSR does not achieve the best SSIM at ×4 on either protocol (WHS-Ortho: 0.9721 vs MsFF-Net's 0.9747; WHS-LAX: 0.8892 vs MsFF-Net's 0.8891). The text mentions 'five out of six metrics at ×4' but this could be stated more prominently to avoid overstating.
- Section 2.2, Eq. (3): The notation switches between superscript s (scale level) and subscript u/v (patch index) without explicit definition of all subscripts. A brief clarification would improve readability.
- Section 3.1: The degradation model (Gaussian σ=1.0 + block averaging) is a standard synthetic approach but does not model realistic MRI acquisition artifacts (coil sensitivity variation, motion, contrast differences). A brief discussion of this limitation would strengthen the paper.
Simulated Author's Rebuttal
We thank the referee for a careful and constructive review. The two major comments are well-taken, and we address each below. In short: (1) we agree that the robustness-to-misalignment claim is not directly validated by the current experiments and will revise the wording and add a controlled perturbation experiment; (2) we agree that the WHS-LAX gains are modest and will add an explicit discussion of this limitation, including qualifying the abstract's language.
read point-by-point responses
-
Referee: Section 1, Contribution (1): The paper claims CFCM enables robust detail transfer under cross-view geometric misalignment, but no experiment introduces residual misalignment. References are resampled from the same volume with known transformations and no inter-acquisition motion. The authors should either add a misalignment experiment or revise the claim.
Authors: We agree with this comment. The referee is correct that the current experiments do not directly test robustness to residual registration error. We want to clarify the intended meaning: the 'misalignment' we refer to is the inherent cross-view geometric discrepancy between orthogonal or LAX views and the target SAX stack — i.e., the reference and target slices are not parallel and their anatomical content is related by a non-trivial geometric transformation, not a simple translation. The coarse-to-fine design is motivated by this cross-view discrepancy, which makes dense correspondence non-trivial even when the inter-acquisition transformation is known. However, we acknowledge that the wording in Contribution (1) — 'robust detail transfer under cross-view geometric misalignment' — can be read as a claim about robustness to residual registration error, which we do not validate. We will take the following steps in the revision: (a) We will add an experiment applying random rigid perturbations (translations and rotations) to the reference views at controlled magnitudes to demonstrate that CFCM degrades gracefully relative to the single-stage matching baseline (McMRSR-style). (b) We will revise the contribution statement to distinguish clearly between (i) the inherent cross-view geometric discrepancy that CFCM is designed to handle and (ii) residual registration error, which is outside the scope of the current method given our stated assumption of pre-aligned inputs (Section 1). We believe this addresses the referee's concern without overstating our claims. revision: yes
-
Referee: Table 1, WHS-LAX rows: Improvements are modest (+0.17 dB PSNR at ×4, +0.19 dB at ×8). Practical significance of sub-0.2 dB gains is questionable. The paper should discuss this limitation explicitly and acknowledge that stronger WHS-Ortho gains may benefit from dense reference coverage less representative of clinical LAX acquisitions. The abstract's 'consistent improvements' overstates the WHS-LAX evidence.
Authors: We agree that the WHS-LAX gains are modest and that this should be discussed more transparently. We will make the following changes: (a) Add an explicit paragraph in Section 3.4 acknowledging that the WHS-LAX improvements are smaller in absolute terms and discussing the likely cause — the LAX protocol provides only three 2-chamber, 3-chamber, and 4-chamber views as references, which cover a small fraction of the target SAX slices, whereas WHS-Ortho uses a dense sagittal reference volume that intersects every axial slice. The denser reference coverage in WHS-Ortho provides more matching opportunities and thus a larger performance gap. (b) Acknowledge that the WHS-Ortho results, while demonstrating the method's potential under favorable reference coverage, may not generalize directly to typical clinical LAX acquisitions where reference views are sparse. (c) Revise the abstract to replace 'consistent improvements' with more precise language, e.g., 'improvements over baselines at ×4 and ×8 upsampling factors, with larger gains under dense reference coverage.' We note that the ×8 WHS-LAX improvement, while small in absolute terms (0.19 dB), is statistically significant (p < 0.001 by Wilcoxon signed-rank test) and consistent across all three metrics (PSNR, SSIM, MSE). We will report effect sizes more carefully and let the reader judge practical significance rather than characterizing the gains as large. We respectfully note that even modest gains at ×8 — the most clinically common through-plane spacing — can be meaningful when no existing method provides a significant improvement, but we agree the paper should not overstate this. revision: yes
Circularity Check
No significant circularity; derivation is self-contained against external benchmarks
full rationale
The paper proposes STRMSR, a reference- and memory-guided through-plane super-resolution framework for cardiac MRI. The central claim—that STRMSR achieves consistent improvements over baselines at 4x and 8x upsampling—is validated against external baselines (Bicubic, MsFF-Net, MINet, MASA, McMRSR) on the public WHS dataset using standard metrics (PSNR, SSIM, MSE) with statistical significance testing (Wilcoxon signed-rank, p<0.001). The method's components (CFCM, PDFA, memory bank) are defined by their own equations (Eqs. 1-7) and validated through an ablation study (Table 2) showing each component's independent contribution. While the paper builds on prior work (McMRSR [13], MASA [15]) and cites them for architectural elements (e.g., cosine similarity matching, FAM/DPRB modules), these citations reference external published work by different author groups, not self-citations. The cosine similarity and normalized correlation operations (Eq. 3) are standard mathematical operations, not definitions that circularly depend on the claimed results. The loss function (Section 2.6) uses standard l1 and k-space losses against ground truth, not against the method's own outputs. The memory bank stores intermediate SR results as additional references for subsequent frames, which is a standard recurrent/autoregressive design pattern—not circular reasoning, as each frame's SR is still supervised against ground truth during training. No step in the derivation chain reduces to its own inputs by construction. The evaluation protocol, while using synthetic degradation on a single dataset, is a standard experimental design choice (correctness risk) rather than a circularity issue. The paper is self-contained against external benchmarks.
Assumptions & free parameters
free parameters (7)
- lambda_rec =
1
- lambda_k =
0.001
- k (block size) =
8
- d (search region) =
13
- p (PDFA patch size) =
8
- r (refinement radii) =
4, 5
- T (memory bank size) =
3 (Ortho), 2/10 (LAX)
assumptions (3)
- domain assumption Input LR stacks are already spatially aligned.
- domain assumption Synthetic degradation (Gaussian pre-filtering + block averaging) accurately models clinical LR acquisition.
- standard math Swin Transformer blocks are effective for feature extraction in MRI SR.
invented entities (1)
-
Memory Bank (FIFO)
independent evidence
Cite this review
Pith. "Pith review of Cardiac MRI Through-Plane Super-Resolution Guided by Reference and Memory." pith.science (2026). https://pith.science/paper/YKJIIPNR
@misc{pith2026260707581,
author = {Pith},
title = {Pith review of: Cardiac MRI Through-Plane Super-Resolution Guided by Reference and Memory},
year = {2026},
howpublished = {\url{https://pith.science/paper/YKJIIPNR}},
note = {Machine review of arXiv:2607.07581}
}
read the original abstract
Clinical cardiac MRI is commonly acquired with high in-plane resolution but coarse through-plane resolution to reduce scan time and accommodate breath-hold and cardiac-motion constraints, which limits 3D analysis and diagnostic accuracy. We propose STRMSR, a reference- and memory-guided through-plane super-resolution (SR) framework that reconstructs high-resolution (HR) cardiac volumes by leveraging HR reference views acquired from the same subject and intermediate SR results as the memory. Our method uses coarse-to-fine contextual matching to establish robust correspondence between low-resolution target and reference/memory images under spatial misalignment. A learnable patch-wise dynamic feature aggregation module predicts content-adaptive mixture weights for each local patch, effectively fusing dynamic information while suppressing unreliable feature transfers. The intermediate SR results stored in the memory bank ensure slice-to-slice consistency for the super-resolved 3D volume. Experiments on the WHS cardiac MRI dataset under two reference protocols, orthogonal-plane views and long-axis chamber views, demonstrate consistent improvements over baselines at 4x and 8x upsampling factors.Code is available at https://github.com/030108ming/STRMSR
Figures
Reviewed July 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.