List of Papers Browse by Subject Areas Author List
Abstract
Volumetric computed tomography (CT) often exhibits severe through-plane anisotropy due to thick-slice acquisition, disrupting structural continuity and degrading downstream lung cancer analysis. We propose TVSRN-V2, a volumetric super-resolution (SR) framework with an anisotropy-aware Through-Plane Attention Block (TAB) that models inter-slice dependencies by factoring attention into orthogonal sagittal and coronal planes. To broaden the degradation distribution during training, we use hybrid pseudo-low-resolution (pseudo-LR) augmentation combining real paired scans with synthetic degradations. Clinical utility is evaluated across pulmonary lobe and bronchi segmentation, histology classification, and survival prediction in 742 patients (539 from three independent external cohorts). TVSRN-V2 achieves PSNR 39.16 dB and SSIM 0.946, and yields an absolute Dice improvement of 0.16 for bronchi over bicubic interpolation on real low-resolution CT. SR features improve histology classification (F1 +0.03–0.07, $p < 0.05$) and internal survival prediction (C-index +0.06, $p=0.02$), with smaller non-significant external prognosis gains.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5469_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BouMar_ThroughPlane_MICCAI2026,
author = { Boubnovski Martell, Marc AND Linton-Reid, Kristofer AND Chen, Mitchell AND Hindocha, Sumeet AND Hunter, Benjamin AND Calzado, Marco A. AND Lee, Richard AND Posma, Joram M. AND Aboagye, Eric O.},
title = { { Through-Plane Attention for Anisotropic CT Super-Resolution Improves Quantitative Lung Cancer Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduced an improved version of super resolution for volumetric CT. The proposed TVSRN-v2 is an asmmetric encoder decoder SR built on top of swin v2.the main contribution the paper claimed is it’s dual attention branch through plane attention block. Another part is a hybrid pseudo low res aug strategy. the evaluation contains common PSNR SSIM, also pulmonary lobe/bronchi segmentation, histology classification, survival prediction on large internal and external cohorts.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.I’m very familiar with the SR works and examined many approaches, this field in general is very challenging because the current benchmarks are extremely high and hard to beat so I welcome the authors effort to improve on top existing methods. The TAB is actual less impressive than hybrid pseudo LR aug. it’s a good approach to mitigate the data size limitation and homogeneity of real paired SR data. 2.the evaluation that improves downstream task is valuable empirical evidence on the importance of SR with a variety of task types. the paper reports consistent imporvements over original TVSRN in several settings, and the segmnetation gain on real low res CT is noticable. 3.the inclusion of extra large external validation is a strength
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.the novelty of the v2 is relatively low compare to previous work. the paper highlighted TAB as central new contribution, but TVSRN already has a TAB inside the decoder’s FIM, with two parallel branches on coronal and sagittal view. the difference is just very minor. 2.the main competition method shouldn’t be TVSRN, it should eb CTHNet which is an improved version of TVSRN and the performance is really good on the CTHNet with several substantial redesign of the architecture. 3.the empirical gains are mixed relative to the strength of the claims. The improvement over the original TVSRN in PSNR is modest, and the downstream gains, while directionally positive, are not uniformly strong across all external prognosis settings. This makes the novelty issue harder to overlook, because the paper is not paired with a clearly empirical jump
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, the paper’s positioning is not clear enough about its relationship to prior works. Because TAB, the RPLHR-CT setting, the asymmetric encoder-decoder formulation, and the general thick-slice-to-downstream-analysis story are all strongly tied to earlier work, the manuscript needs much clearer attribution and framing.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The response is constructive and honest, the main concern is still the central framework novelty is weak. the response narrows the difference to adaptive gating, swin v2 modernization, and hybrid pseudo-lr training. those are good engineering efforts, just it’s very incremental. that said, it’s still is a valuable work because of the clinical validation, but not around TAB novelty.
Review #2
- Please describe the contribution of the paper
This paper proposes a transformer-based volumetric super-resolution framework (TVSRN-V2) for anisotropic CT, introducing a through-plane attention mechanism to explicitly model inter-slice dependencies and demonstrating its impact on downstream lung cancer analysis tasks.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The problem is important, common in clinical practice, and well-motivated.
- The method is well-designed with a clear focus on through-plane modeling.
- The evaluation is comprehensive, including multi-center validation and multiple downstream tasks.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The comparisons are not fully sufficient to establish the necessity and uniqueness of the proposed design, lacking stronger 3D or inter-slice modeling baselines.
- The paper provides limited analysis on why downstream improvements are achieved beyond reporting performance gains.
- The presentation can be improved, e.g., Figure 1 is unclear and Table 3 is difficult to read due to small font.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The score is based on the above strengths and weaknesses.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Although the response clarifies some architectural differences from TVSRN, it does not adequately address my core concern regarding the necessity of the proposed design, especially why TAB is preferable to other plausible 3D or inter-slice modeling strategies and how it leads to the reported downstream improvements.
Review #3
- Please describe the contribution of the paper
The main contribution of this work is the introduction of the TAB module, which restores structural continuity by employing two orthogonal attention pathways. A secondary contribution is the use of pseudo-low-resolution (pseudo-LR) augmentations during training. Combined with real paired data, this strategy aims to improve generalization across scanners while minimizing domain gaps. Finally, the approach is evaluated extensively on multiple datasets, including ablation studies, to assess both performance and potential clinical utility.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
(S1) The paper provides an extensive evaluation of clinical utility across multiple downstream tasks, covering five acquisition protocols and three distinct tasks. The results of the proposed model are promising compared to the two alternative approaches evaluated.
(S2) The authors address the scarcity of high-quality paired super-resolution (SR) datasets by incorporating pseudo-low-resolution (pseudo-LR) images during training, which is a well-motivated strategy to improve generalization.
(S3) Statistical testing is performed, and confidence intervals are provided, supporting the reliability of the reported results.
(S4) The manuscript includes a discussion of the limitations of the proposed approach, providing context for the scope and applicability of the method.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
(W1) The related work is limited, with only a single reference ([13]) cited for existing CT super-resolution (SR) approaches. Several recent methods, including publications from the last few years at MICCAI and other venues, are not discussed or compared. I recommend performing a focused literature search for recent CT SR approaches and citing them appropriately. Including these references would provide a more complete context and support claims made in, for example, the final sentences of Subsection 2.2. (W2) While downstream task performance is reported, a qualitative assessment of the SR outputs is missing. No visual comparisons between SR and high-resolution (HR) images are provided, making it difficult to evaluate structural consistency and the potential for hallucinated features.
(W3) The manuscript does not report inference speed, which is an important factor for assessing clinical feasibility in real-world settings.
(W4) No systematic or controlled hyperparameter optimization is reported, limiting confidence in whether the results reflect an optimized configuration.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
-
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
(C1) Abbreviations are not consistently introduced and used throughout the manuscript. Each abbreviation, including common ones such as CT and LR, should be defined upon first use in both the abstract and main text, and then used consistently thereafter.
(C2) Several variables are not properly defined, which reduces clarity and makes some equations harder to follow.
(C3) Some statements and components lack references (e.g., Section 2.2). Each method, claim, or component should be properly cited.
(C4) Headings should not appear consecutively without intervening text (e.g., Section 2 followed directly by Subsection 2.1). It would also improve readability to include a brief introductory sentence at the start of Subsection 2.1. (C5) There are undesirable line breaks in several places (e.g., page 3: “[…] LUNA16 dataset (<= LINEBREAK 3 mm) […]” and page 5 before “(C-Index)”). These could be corrected using non-breaking spaces to improve formatting and readability.
(C6) Variables in the text and equations should be written in italics (e.g., p in the abstract).
(C7) References should be ordered numerically when multiple are cited together (e.g., page 1 [1, 2] instead of [2, 1]).
(C8) Equations should be properly integrated into the text with correct punctuation and formatting to improve readability.
(C9) The number of GPUs used for training is not explicitly stated and should be reported.
(C10) The orange text in Figure 1 is difficult to read; adjusting the color or contrast would improve visibility.
(C11) As a suggestion, considering the Gradient SSIM (G-SSIM) metric may provide a more sensitive evaluation for SR tasks in addition to standard SSIM.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper introduces a novel architectural component within an existing model and validates it through extensive ablation studies. The reported results are statistically significant and indicate promising performance. However, several recent related works are not discussed or compared, which limits the contextualization of the contribution. In addition, there are stylistic and formatting issues that should be addressed to improve clarity and presentation.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors addressed most of the weaknesses identified in my original review, with the exception of (W4), which remains unresolved. In particular, the rebuttal provided additional clarification regarding the TAB module, the role of the hybrid pseudo-LR strategy, and the planned qualitative evaluation of structural consistency. Assuming that the improvements and corrections outlined in the rebuttal, including the addressed comments, are incorporated into the final manuscript, I would support acceptance of the paper.
Author Feedback
Relationship to TVSRN We agree that TVSRN-V2 is an evolutionary rather than paradigmatic advance. Our contribution is a clinically oriented redesign targeting through-plane structural recovery in anisotropic CT, combined with a robust training strategy and evaluated across multiple real-world tasks and institutions.
TAB Design and Distinction from TVSRN (R2) The original TVSRN uses Feature Interaction Modules (FIMs) for cross-scale fusion, with coronal and sagittal interactions appearing only during local decoder refinement. TAB, described in Section 2.2, is architecturally distinct: it reshapes intermediate features into orthogonal (D,H) and (D,W) representations, applies shifted-window self-attention independently in each plane, and fuses the results via adaptive learned gating:
F_TAB = σ(W_g ⋅ [F_sag; F_cor]) ⊙ F_sag + (1 − σ(W_g ⋅ [F_sag; F_cor])) ⊙ F_cor
This differs from implicit inter-slice aggregation by decomposing volumetric attention into orthogonal pathways before adaptively recombining them. The ablation in Section 3.1 (Table 1) supports this design: TAB removal produced the largest single-component degradation (−1.34 dB PSNR, −0.026 SSIM), and the largest downstream gains occurred in anatomies most sensitive to through-plane averaging — bronchi and lobes.
CTHNet and Related Work (R1) We thank R1 for identifying the convolutional-transformer hybrid of Yu et al. (npj Digital Medicine, 2024) as a relevant baseline. The T-CTH block similarly processes coronal and sagittal pathways in parallel, but fuses them via element-wise summation. TVSRN-V2 uses adaptive channel-wise gating instead, allowing the network to weight orthogonal evidence dynamically according to local anatomy. Importantly, Yu et al. identify training data homogeneity as a primary limitation of their framework — a bottleneck our hybrid pseudo-LR strategy is designed to address directly. Per MICCAI guidelines, we cannot introduce new empirical comparisons during rebuttal, but we will revise the related work section to contextualize CTHNet fully and address the runtime considerations raised by R3. Hybrid Pseudo-LR Augmentation (R1) As described in Section 2.1, the hybrid pseudo-LR strategy is a core component of the framework, combining real paired RPLHR-CT data with synthetic degradations spanning heterogeneous slice thicknesses to improve robustness across scanners and acquisition protocols. We will revise the manuscript to foreground this contribution more explicitly alongside the TAB design, clarifying that the two together constitute the primary methodological advance.
Structural Consistency and Qualitative Evaluation (R3, AC) Given the text-only rebuttal format, we cannot include figures here. The final manuscript will add representative side-by-side comparisons — Bicubic, TVSRN, TVSRN-V2, and HR target — with emphasis on continuous airway and vascular structures where through-plane discontinuity is most consequential. Quantitatively, the bronchi Dice improvement on real low-resolution CT (0.508 to 0.670, Table 3) provides the strongest available evidence of structural recovery: bronchi topology is highly sensitive to through-plane averaging and serves as a critical landmark in tumor assessment. We recognize this metric addresses airway fidelity specifically; the final visualizations will extend coverage to vascular and soft-tissue boundaries to more fully address concerns about structural hallucination.
Presentation and Notation (R1, R2, R3) We will address all noted issues regarding notation, citations, and figure readability in the final version.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper had mixed reviews that would be improved by a rebuttal. The reviewers noted a large external validation, multiple downstream tasks, and statistical testing as positives in the study, but the main uncertainty was whether the claimed architectural novelty is real relative to prior TVSRN/CTHNet-style methods. The paper presents TAB as the central contribution, but the original TVSRN already appears to contain a related through-plane attention block inside its decoder/FIM with parallel coronal and sagittal branches. If true, the architectural novelty is minor. Notes were also made that recent CT SR literature was missing from the related work and comparisons. The rebuttal should focus on: 1) explaining precisely how TAB differs from prior TVSRN/FIM and whether CTHNet is included, unavailable, or not comparable, 2) cite and contextualize recent CT SR methods, especially those with 3D or inter-slice modeling, and 3) include qualitative SR/HR comparisons, if available.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors were straightforward and clear in their rebuttal, but there are still remaining concerns that were not (or could not) be addressed at this stage. In particular, there is a need to evaluate stronger baselines and to compare the TAB module against other similar methodologies. While this remains unaddressed, the reviewers highlight the careful rebuttal of their concerns and the powerful clinical validation. In the end, the contribution is solid and well-positioned in clinical validation, if incremental. I recommend acceptance of this article, but in remains borderline.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The key contribution is the component introduced for super resolution – novelty of TAB component must be clarified and explained in terms of other prior works as to how it improves super resolution accuracy. All reviewers were concerned regarding the engineering focus of the innovation so it is important to really clarify and frame the contribution in relation to prior works better.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Concerns regarding the technical novelty and justification of the proposed design were not fully addressed, but reviewers agreed that it is a valuable work, provided the improvements are incorporated into the final version.