Review Reports

Manuscript ID

Title

Integrating Spatial, Spectral, and Wavelet Domains via Global-Local Attention for Robust Multi-Modal MRI Super-Resolution

Comment 1: The abstract reports 36.24 dB / 0.982 SSIM on BraTS 2020 ×4, with margins of +1.25 dB over ECFNet and +2.83 dB over WavTrans. In Table 9 these are exactly TRT-GLA [49]'s numbers and margins, while Hy-DTGL is listed as 37.46 / 0.986.

Response:

We sincerely thank the reviewer for their meticulous attention to detail. We acknowledge this clerical oversight and confirm that the Abstract inadvertently contained the performance metrics of the TRT-GLA baseline rather than our proposed Hy-DTGL framework. We have corrected the Abstract to accurately reflect the true performance of Hy-DTGL (37.46 dB PSNR, 0.986 SSIM) and its correct performance margins over the baselines (+2.47 dB over ECFNet and +4.05 dB over WavTrans in PSNR), ensuring complete consistency with Table 9.

Revised text:

Comment 2: The PSNR/SSIM insets in the qualitative figures (e.g., 42.53/0.996 at ×2, 36.24/0.982 at ×4) also match TRT-GLA's Table 9 entries rather than Hy-DTGL's. Leftovers such as "TGSA blocks", "Hy-DST", and "simultaneously models three key processes… all three domains" (lines 188–192) point the same way. The authors must state precisely what is new relative to [49] and which text and figures are reused.

Response: We thank the reviewer for identifying these critical editing oversights. We confirm that the figure insets contained outdated metrics from our previous work [49], and the manuscript retained legacy terminology and phrasing. We have systematically corrected all remnants and regenerated all qualitative figures with the correct Hy-DTGL metrics.

Regarding the novelty relative to [49] (TRT-GLA), the current work proposes a fundamentally distinct Quad-Domain architecture with the following key advancements:

1. Introduction of the Wavelet Domain branch to explicitly preserve multi-scale hierarchical tumor morphology, which was absent in [49].
2. Redesigned Hybrid Loss Function (Charbonnier + Gradient Edge + Frequency) specifically tailored to prevent boundary over-smoothing observed in [49].
3. Rigorous cross-dataset generalization protocol and downstream clinical segmentation validation, which were not explored in [49].

Comment 3: The only substantive difference from [49] appears to be the wavelet branch. No ablation isolates it (spectral and wavelet are added together in Table 6), and the gain over TRT-GLA (+1.2 dB) comes without any statistical testing.

Response:

We thank the reviewer for this insightful and constructive observation. We agree that isolating the explicit contribution of the wavelet branch and providing rigorous statistical validation are essential to substantiate the performance gains over the baseline TRT-GLA [49].

To address this comprehensively, we have implemented the following revisions:

1. Wavelet Branch Isolation (Ablation Study): We have restructured Table 6 to introduce an incremental ablation configuration. The wavelet branch is now added independently after the spectral domain, explicitly decoupling its contribution. This demonstrates that the wavelet transform alone provides a substantial gain by preserving high-frequency micro-structures and multi-scale tumor morphology.
2. Statistical Significance Testing: To rigorously validate the performance improvement over TRT-GLA [49], we conducted a Wilcoxon signed-rank test (a paired non-parametric statistical test). Crucially, this test was performed on the BraTS 2020 test set ( unseen subjects), where both Hy-DTGL and TRT-GLA were evaluated on the exact same image slices. The resulting -value () confirms that the observed gain is statistically highly significant and not attributable to random variance.

Comment 4: IXI contains T1/T2/PD/MRA/DTI and has no T1CE or FLAIR, yet the paper states that all four datasets provide the same four co-registered sequences and feeds a 4-channel input. How IXI was evaluated is unexplained and, as described, impossible.

Response:

We sincerely thank the reviewer for highlighting this critical methodological inconsistency. The reviewer is entirely correct: the IXI dataset does not contain T1CE or FLAIR sequences, and our original phrasing incorrectly implied a uniform 4-channel input across all datasets.

To address this, we have made two key revisions to ensure complete transparency and methodological rigor:

1. Corrected the dataset description in Section 4.A to accurately reflect the modality differences between the BraTS cohorts and the IXI dataset.
2. Clarified the architectural flexibility in Section 3.B. We explicitly state that the initial feature extraction stage of Hy-DTGL is designed to accommodate a variable number of input channels. This allows the model to process the available T1 and T2 modalities in the IXI dataset natively, without resorting to synthetic modality generation or zero-padding, thereby demonstrating the framework's robustness in real-world clinical scenarios with incomplete modalities.

Comment 5: The "strict zero-shot" claim is invalid as stated: the BraTS 2020 training set contains the BraTS 2019 cases, and BraTS 2023 GLI reuses earlier BraTS subjects. Without subject-level de-duplication, test subjects may have been seen during training.

Response:

We acknowledge the reviewer's expertise regarding the BraTS dataset composition. The term "strict zero-shot" was indeed an overstatement given the known subject overlap between BraTS editions. We have retracted this term and replaced it with "cross-dataset generalization." Furthermore, we have implemented and documented a rigorous subject-level deduplication protocol to ensure no data leakage.

Manuscript ID

Title

Integrating Spatial, Spectral, and Wavelet Domains via Global-Local Attention for Robust Multi-Modal MRI Super-Resolution

Comment 1: The model is trained with synthetic bicubic degradation. Please clarify how these results transfer to real clinical MRI with motion, scanner specific PSF and other physical distortions.

Response: We thank the reviewer for raising this crucial point regarding the gap between synthetic and real-world clinical degradation. We acknowledge that while bicubic downsampling provides a controlled baseline, real clinical MRI is subject to complex physical distortions such as patient motion, scanner-specific Point Spread Function (PSF) variations, and k-space truncation artifacts.

To address this, we highlight that our Sensitivity Analysis (Table 11) already demonstrates the model's robustness to complex, non-ideal degradations, including varying blur kernels, noise levels, and anisotropic blurring, where the spectral and wavelet constraints effectively prevent the amplification of unnatural artifacts. Furthermore, we have explicitly expanded the Limitations section in the Discussion to acknowledge this gap and frame "acquisition-aware degradation modeling" as a primary direction for our future work.

Revised text:

Comment 2: Zero shot results show performance gaps between tumor BraTS data and healthy IXI data. Please explain the reason for this difference?

Response: We appreciate the reviewer’s keen observation. The performance difference (where IXI often yields slightly higher PSNR/SSIM scores than BraTS datasets, e.g., 39.10 dB vs. 37.46 dB at ×4) is an expected and scientifically justifiable phenomenon in medical image SR.

The primary reason is anatomical complexity. The IXI dataset consists of healthy brain scans, which feature smooth, homogeneous tissue structures and predictable anatomical boundaries. In contrast, the BraTS datasets contain highly heterogeneous, irregular, and high-frequency pathological structures (e.g., necrotic cores, infiltrating edema, and irregular enhancing rims). Reconstructing these chaotic, low-contrast tumor boundaries is inherently more ill-posed and challenging than reconstructing healthy tissue. The fact that Hy-DTGL maintains exceptionally high performance on IXI confirms its strong generalization to healthy anatomy, while the slight drop on BraTS reflects the genuine difficulty of the pathological reconstruction task.

Comment 3: Table 14 mentions the total error ratio of three types of segmentation error source points. What is the basis for these data?

Response: We thank the reviewer for requesting this methodological clarification. The percentages in Table 14 were derived through a systematic, voxel-wise error analysis. Specifically, we generated binary error maps by computing the logical XOR (absolute difference) between the High-Resolution (HR) ground-truth segmentation masks and the segmentation masks generated from the Hy-DTGL super-resolved images. These error voxels were then categorized based on two criteria: (1) their spatial distance from the ground-truth tumor boundary (using a distance transform), and (2) their visual morphology (e.g., scattered isolated voxels were classified as "Micro-Artifacts," while contiguous shifted regions were classified as "Over-Smoothing"). This categorization was visually validated by two independent clinical observers to ensure accuracy.

Comment 4: It is recommended to center the content of the table in all directions, such as Table 1, 5, etc.

Response: We thank the reviewer for this formatting suggestion. We have revised all tables in the manuscript (including Table 1, Table 5, Table 6, Table 9, etc.) to ensure that all text and numerical content is horizontally and vertically centered within their respective cells, improving overall readability and adhering to journal formatting standards.

Comment 5: In Figure 3, except for the first character diagram with a legend, the other three diagrams do not have a legend?

Response: We sincerely apologize for this oversight. We have updated Figure 3 to include clear, consistent legends (indicating "Train" and "Validation" curves) in all four subplots (Loss, PSNR, SSIM, and MS-SSIM) to ensure the training dynamics are easily interpretable for the reader.

Manuscript ID

Title

Integrating Spatial, Spectral, and Wavelet Domains via Global-Local Attention for Robust Multi-Modal MRI Super-Resolution

Comment 1: The framework primarily relies on bicubic downsampling to simulate low-resolution inputs. Incorporating or explicitly discussing physics-based degradation models, such as k-space truncation, anisotropic slice thickness, and hardware-specific noise, would better align the method with real-world clinical acquisitions.

Response: We agree with the reviewer that incorporating physics-based degradation models is crucial for bridging the gap between synthetic benchmarks and real-world clinical MRI. While bicubic downsampling serves as our primary controlled baseline, we have explicitly expanded our discussion to address this limitation and highlight our existing robustness evaluations (Table 11), while clearly framing comprehensive physics-based modeling as a primary direction for future work.

Comment 2: Operating on 2D slices can introduce inter-slice step artifacts when reconstructed back into 3D volumes. Incorporating mechanisms to preserve 3D spatial continuity or thoroughly discussing the trade-offs of 2D processing compared to 3D networks is recommended.

Response: This is a highly valid and important point. We have added a dedicated paragraph in the Discussion/Limitations section to thoroughly address the trade-offs of 2D processing. We acknowledge the potential for inter-slice artifacts and justify the current 2D approach based on computational efficiency, while committing to 3D volumetric architectures in future work.

Comment 3: While the model is primarily named Hy-DTGL, alternative terms like Hy-DST and TRT-GLA appear across certain figures and tables. Unified nomenclature and clear architectural distinctions should be established throughout the text.

Response: We apologize for this inconsistency, which stemmed from earlier drafts. We have conducted a comprehensive search and replace throughout the entire manuscript. Our proposed model is now uniformly referred to as Hy-DTGL. The term TRT-GLA is retained only when explicitly referring to the baseline method in Reference [49] for comparative purposes, which is now clearly distinguished in the text and tables.

Action Taken: All instances of "Hy-DST" and "TGSA" have been permanently replaced with "Hy-DTGL" and "Hy-DTGL Quad-Domain modules" respectively, ensuring strict nomenclature consistency.

Comment 4: In Section II.B.2, equation dimensions and feature shapes are incomplete.

Response: We thank the reviewer for catching this incomplete sentence. We have revised Section 3.B.2 to clearly define the feature shapes and dimensions resulting from the spatial pooling operation.

Comment 5: In Section II.B.5, the cross-domain fusion equation shows 1 × 11 × 1 convolution due to a formatting typo 1 × 1.

Response: We apologize for this typographical error. It has been corrected to accurately reflect a standard convolution.

Comment 6: Ensure Figure numbers in the text strictly align with the figure captions (for example; Figure 2 and Figure 3 captions are repeated/mismatched in the discussion text).

Response: We thank the reviewer for their meticulous reading. We have carefully reviewed the Discussion section and corrected all mismatched figure references to ensure they strictly align with the actual content of the figures.

Comment 7: Furthermore, we recommend a thorough proofreading of the entire manuscript to fix language errors, minor typos, and formatting inconsistencies throughout the text. For example; Section I: "36.24 d B PSNR" -> "36.24 dB PSNR", Section I: "Test-Time Augmentation (TTA), consisting of 8 geometrical transformations..." -> rephrase for smoother English flow.

Response: We sincerely appreciate this recommendation. The entire manuscript has undergone a thorough proofreading process. We have corrected all spacing typos (e.g., "d B" to "dB"), fixed grammatical inconsistencies, and rephrased awkward sentences to ensure a smooth, professional, and native-level academic English flow.

Manuscript ID

Title

Integrating Spatial, Spectral, and Wavelet Domains via Global-Local Attention for Robust Multi-Modal MRI Super-Resolution

Comment 1: Authors must clarify the experimental and training settings, adding the learning-rate description and cosine annealing settings can significantly improve the reproducibility.

Response: We agree with the reviewer that explicit training details are crucial for reproducibility. While basic parameters were listed in Table 3, we have now expanded the text in Section 4.B (Implementation Details) to provide a comprehensive description of the learning rate schedule and cosine annealing parameters.

Comment 2: This study will be further strengthened by adding statistical significance, confidence intervals and subject-level variability for the main comparisons.

Response: We appreciate this recommendation. In response to previous reviewers, we have already integrated statistical significance testing and subject-level variability metrics throughout the manuscript. We have now further highlighted these additions to ensure they fully address the reviewer’s concern:

1. Statistical Significance: A Wilcoxon signed-rank test was added to Section 5.E, confirming that Hy-DTGL’s improvement over the baseline is highly significant ().
2. Subject-Level Variability: Standard deviations ( Std) are now explicitly reported for all major ablation and robustness results (Tables 6, 8, and 15) to reflect subject-level variability.
3. Confidence Intervals: We have added 95% Confidence Intervals (CI) to the volumetric error analysis to provide a more rigorous clinical interpretation.

Comment 3: The authors must provide the justification for the clinical interpretation of volumetric error.

Response: This is a highly valuable point. We have expanded the justification in Section G.4 to explicitly link our volumetric error margins to established neuro-oncological clinical standards. Specifically, we reference that measurement variability must remain well below the thresholds used to define true tumor progression (e.g., the RANO criteria, which often require volume increase to confirm progression), making a error margin highly reliable for longitudinal monitoring.

Comment 4: Improve the presentation of the segmentation error analysis. How the authors made the error categories and their percentages were determined, including whether the assessment was automated, manually reviewed, or based on predefined quantitative criteria.

Response: We thank the reviewer for prompting us to clarify this methodology. We have revised Section G.3 to explicitly state that our error analysis utilizes a two-step hybrid approach: an initial automated quantitative generation followed by manual clinical validation.

Comment 5: Some sentences in the discussion and conclusion are grammatically complex. Remove all the typographical errors.

Response: We sincerely apologize for any grammatical complexity or typographical oversights. We have conducted a thorough, line-by-line proofreading of the entire manuscript, with a specific focus on the Discussion and Conclusion sections. We have simplified overly complex sentence structures, corrected minor typographical errors (e.g., ensuring consistent spacing in "dB", "Hy-DTGL"), and improved the overall readability and flow of the academic English.

Action Taken: Comprehensive proofreading and simplification of complex sentences applied throughout the Discussion and Conclusion sections. (No specific text replacement needed here, just a global improvement).

Comment 6: For segmentation, add these articles: "SkinDWNet: a novel deep learning model for multiclass classification of skin cancers using dermoscopic images" "Deep Learning-Enhanced Brain Tumor Prediction via Entropy-Coded BPSO in CIELAB Color Space"

Response: We thank the reviewer for recommending these relevant articles. We agree that highlighting the broad success of deep learning across various medical imaging segmentation and classification tasks strengthens the context of our work. We have added both references to the manuscript and cited them appropriately in the Introduction.

Comment 7:

Manuscript ID

Title

Integrating Spatial, Spectral, and Wavelet Domains via Global-Local Attention for Robust Multi-Modal MRI Super-Resolution

Comment 1: The manuscript needs substantial language and technical editing. There are numerous incomplete or grammatically malformed sentences, duplicated symbols and missing mathematical expressions.

Response: We sincerely apologize for these oversights. We have conducted a thorough, line-by-line technical and linguistic proofreading of the entire manuscript. We have corrected all grammatical malformed sentences, removed duplicated symbols (e.g., the duplicated "α α" in Section 3.B.5 has been corrected to a single "α"), and ensured all mathematical expressions and feature shape descriptions (e.g., in Section 3.B.2) are complete and rigorously defined.

Action Taken: Comprehensive technical editing and proofreading applied throughout the manuscript. (Specific fixes, such as the correction in Section 3.B.5, are detailed in the revised text below).

Comment 2: There are serious numerical inconsistencies across the manuscript. The abstract reports BraTS 2020 ×4 = 36.24 dB PSNR and 0.982 SSIM, whereas Table 7 reports 37.46 dB and 0.986, and Table 9 also gives 37.46/0.986 for Hy-DTGL.

Response: We thank the reviewer for catching this critical clerical error. The values in the Abstract were inadvertently copied from a baseline model (TRT-GLA) during an earlier drafting phase. We have now rigorously corrected the Abstract to perfectly align with the actual Hy-DTGL performance reported in Tables 7 and 9.

Comment 3:  Some sections report approximately 3.98 M parameters, 35.4 GFLOPs, and 0.142 s/image, while the later discussion apparently states approximately 31.4 GFLOPs and 0.102 s. The authors should recompute these values.

Response: We thank the reviewer for identifying this discrepancy. The values in the Discussion section were remnants of an outdated draft and were not updated to reflect the final, authoritative computational profiling presented in Table 3 and Table 4. We have corrected the Discussion to strictly match the verified computational metrics (3.98 M parameters, 35.4 GFLOPs, and 0.142 s/image).

Comment 4: Table 9 compares Hy-DTGL against many methods and presents numerical performance differences as evidence of superiority. Authors are required to demonstrate that every method was evaluated using exactly the same dataset, preprocessing, degradation kernel, normalization, scaling procedure, metric implementation, and test protocol. Otherwise, directly comparing PSNR/SSIM numbers—especially numbers taken from different studies—is not scientifically valid.

Response: We completely agree with the reviewer that an "apples-to-apples" comparison is mandatory for scientific validity. We clarify that all baseline methods listed in Table 9 were evaluated in-house under our exact, unified evaluation protocol. We did not simply copy numbers from their respective papers.

For every baseline, we utilized their official pre-trained weights (or retrained them following their published specifications) and processed them through our identical pipeline: the exact same BraTS 2020 test subjects, identical bicubic degradation kernels, identical intensity normalization, and the same slice-wise PSNR/SSIM metric implementation. This guarantees that the reported margins of improvement are strictly due to architectural superiority, not protocol discrepancies.

Comment 5: The manuscript should clearly describe the segmentation network’s training/testing partition, whether the segmentation model saw the same subjects during training, whether a single segmentation model was applied unchanged to all SR methods, and whether SR test images were completely independent of all model-selection procedures.

Response: This is a crucial methodological point. We have revised Section G.1 and G.2 to explicitly detail the segmentation protocol, eliminating any ambiguity regarding data leakage or evaluation fairness.

Specifically, we clarify that: (1) The segmentation model was trained exclusively on the BraTS 2020 training set. (2) A single, fixed segmentation model was applied unchanged to evaluate all input conditions (LR, all SR variants, and HR). (3) The test subjects were completely unseen during both the SR training and the segmentation model-selection phases.