List of Papers Browse by Subject Areas Author List
Abstract
Susceptibility-Weighted Imaging (SWI) is essential for detecting Cerebral Microbleeds (CMBs), a key biomarker in neurodegenerative diseases. However, clinical SWI scans are often acquired with thick slices within clinically acceptable scan durations, resulting in severe anisotropy that degrades through-plane resolution and leads to missed small lesions. To address this, this paper proposes Cubic Structure-Aware Neural Radiance Fields (CSANeRF), a zero-shot Super-Resolution (SR) framework that integrates NeRF with Transformer architectures for high-fidelity SWI SR reconstruction. To preserve the critical biomarkers in SWI, we rethink the construction of implicit radiance fields by incorporating explicit structural priors from local 3D neighborhoods. The core innovation lies in a cube-based spatial modeling strategy. CSANeRF is validated on two independent, expert-annotated SWI datasets and consistently outperforms state-of-the-art 3D SR approaches in both visual quality and quantitative metrics. Critically, when employed as a preprocessing step for CMB segmentation, CSANeRF reconstructions substantially improve segmentation accuracy, achieving performance closest to the HR ground truth and surpassing all baseline methods.This finding underscores its potential to enhance diagnostic reliability in neurodegenerative disease evaluation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0708_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{BiHon_Cubic_MICCAI2026,
author = { Bi, Hongwei AND Zhang, Rong AND Guo, Lijun AND Liu, Fengchun AND Xie, Guomin AND Lv, Zhongyue AND Wu, Xiping},
title = { { Cubic Structure-Aware Neural Radiance Fields for Zero-Shot Super-Resolution of Susceptibility-Weighted Imaging } },
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
This paper presents Cubic Structure-Aware Neural Radiance Fields (CSANeRF) for zero-shot Super-Resolution (SR) on clinical SWI scans, bridging the gap of thick slice problem.
- 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.
- Key backgrounds on CMB lesions and SWI scans are well-written.
- Extensive ablation studies were performed, covering lattice sampling numbers, segmented attention, axis traversal strategies, and relative position bias.
- Downstream CMB segmentation was performed, further highlighting the effectiveness of the proposed 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.
- Current evaluation is limited. Other than NeRF, SR can be achieved using GAN or Diffusion-based methods [1]. However, only NeRF-based methods were compared in this paper. As far as I’m concerned, NeRF is not conventional for SR tasks.
- Evaluations were limited to datasets with small number of subjects (n=20 for HKCU-SWI-CMB and n=34 for private dataset). While I acknowledge the difficulty of collecting SWI scans, to see if the proposed method is clinically applicable, authors must validate their method on held-out dataset or use publicly available SWI scans (e. g. , ADNI, AIBL as described in [2]). Neither of these validation strategies was employed.
- There is no information regarding data splitting, which raises concerns about reproducibility. Data splitting information was only reported in Section 3: Applicability to Downstream Tasks. Then how did authors split the data used in Table 1?
- Following [2], further validation should have been performed using GRE MRI rather than MSD-T1.Some GRE scans are publicly available in ADNI and used in [2]. Since T1 MRI resolution is already much better than that of SWI, applying the proposed method to T1 is less meaningful. Given that the key motivation is addressing large slice thickness in imaging protocols, using GRE is a much more appropriate fit for this problem. [1] Z. Ji et al. “Deep learning-based magnetic resonance image super-resolution: a survey” Neural Computing and Applications, 2024. [2] T. Hassanzadeh et al. “A robust deep learning framework for cerebral microbleeds recognition in GRE and SWI MRI” NeuroImage: Clinical, 2025.
- 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 provide sufficient information for 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the methodology is sound, how authors curated the test set remains unknown. This limits the clinical applicability and reproducibility of the method in application studies.
- 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.
I admit there were misunderstandings in the initial review. All of my concerns were successfully addressed.
Review #2
- Please describe the contribution of the paper
The paper proposes a structure-aware, zero-shot NeRF-based method for SWI super-resolution that uses cube-based sampling and Transformer modeling to improve preservation of fine anatomical structures.
- 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.The paper addresses a clinically relevant problem (SWI anisotropy and small lesion detection). 2.The proposed zero-shot method does not require HR training data, making the method more practical and easier to deploy. 3.The proposed method combines cube-based sampling with a Transformer-enhanced NeRF to better capture structural information. 4.The paper includes a fairly complete experimental design, covering both SR reconstruction quality and downstream segmentation performance.
- 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 paper states LR data are generated by z-axis downsampling, but no details on slice profile or noise modeling. 2.Important imaging details (e.g., acquisition parameters and voxel size) are not provided. 3.The method is evaluated exclusively on synthetically downsampled data, without validation on real low-resolution acquisitions.
- 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?
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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant problem and proposes a technically sound solution that is sufficiently novel. The experimental design is also fairly complete, including both reconstruction quality evaluation and downstream segmentation validation across multiple datasets, which supports the effectiveness of the method. While there are some limitations, such as the lack of real LR validation and incomplete imaging details, the overall contribution and empirical results justify acceptance.
- 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.
As the author proposed zero-shot super-resolution method, it is fair for them to only compare with other zero-shot methods
Review #3
- Please describe the contribution of the paper
This work improves upon CuNeRF by incorporating spatial correlations among sampling points rather than processing each point independently. It achieves state-of-the-art performance and addresses an important clinical problem, i.e., zero-shot super-resolution of susceptibility-weighted imaging.
- 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.Zero-shot. No paired HR data needed — trains on the patient’s own LR volume only. 2.Spatial correlation. The proposed method replaces point-independent MLP, enabling information exchange among neighboring sampling points. 3.Comprehensive comparison. Multiple baselines, three datasets, three SR scales — thorough and convincing. 4.Good visualization. Axial, coronal, and sagittal views clearly show sharper details. 5.Clinical significance. The clinical problem is clearly stated and valid.
- 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.Limited baselines. No comparison with supervised SR methods or diffusion-based approaches. Only compares against interpolation and NeRF-family methods. 2.Slow inference. Zero-shot means training a separate model per patient. The paper doesn’t discuss inference time, but training 25000 iterations per volume is not trivial for clinical deployment.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is solid. The idea makes sense; the reason for improvement is clearly stated; the experiment is sufficient. The clinical significance is strong.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 my concerns in the rebuttal about comparison methods and inference speed.
Author Feedback
We thank the reviewers and meta-reviewer for their constructive comments. Their concerns are addressed below.
To Rev #1 W1(Why use NeRF and lack of comparisons to GAN/Diffusion methods): NeRF provides a zero-shot, subject-specific framework for 3D SR. Our CSANeRF is built upon this paradigm to address the practical challenge of scarce clinical SWI data. By training solely on a single subject’s LR volume without external supervision, our method is well suited to this data-limited setting. In contrast, GAN- and diffusion-based SR methods typically rely on large-scale supervised training, making them less suitable for the data-scarce scenarios considered in this study. Therefore, we compare CSANeRF with NeRF-based zero-shot baselines and interpolation methods, enabling a fair and controlled evaluation.
W2 (Limited datasets) and W3(Data splitting): A key advantage of NeRF-based frameworks lies in their self-supervised, zero-shot formulation. The model is trained exclusively on a single subject’s LR volume, with the HR data completely withheld until final evaluation. Consequently, SR training does not rely on a conventional cross-subject train/test split or large-scale external datasets. The results in Table 1 report the average metrics across all subjects to validate the method’s performance. Regarding data splitting mentioned in Section 3, it pertains specifically to our downstream CMB segmentation experiments.
W4(GRE vs. T1 validation): Our dataset selection was primarily driven by clinical relevance. Both SWI datasets used in our experiments were acquired from stroke patients, closely aligning with real clinical scenarios where SWI is the standard sequence due to its high sensitivity to CMBs. T1-weighted MRI has relatively high resolution. We used it as a benchmark for high-fidelity reconstruction to verify that CSANeRF can preserve fine anatomical structures without introducing artifacts. This further demonstrates the robustness of our method across different MRI contrasts.
To Rev #2 W1(LR degradation modeling): You are correct that we adopted a simplified decimation strategy for downsampling along the z-axis, whereas real clinical MRI acquisition involves complex physical degradation. However, we adhered to the established protocol used in CuNeRF. It ensures a fair, controlled, and reproducible comparison with prior work.
W2(Acquisition parameters): Both SWI datasets consist of stroke patients. The private dataset parameters are TR = 31 ms, TE = 7.2 ms, and voxel spacing = 0.6 × 0.6 × 1.5 mm³; the HKCU-SWI-CMB dataset parameters are TR = 17 ms, TE = 24 ms, and voxel spacing = 0.45 × 0.45 × 1.0 mm³. These details will be added to the final manuscript.
W3(Validation on real LR acquisitions): We agree that validation on real LR acquisitions is ideal. However, in clinical practice, SWI is typically acquired at high resolution for reliable detection of subtle lesions such as CMBs, and real LR SWI volumes are rarely available for research. Nevertheless, validating our method on real LR MRI acquisitions from different imaging sequences and further exploring its potential to reduce patients’ multi-sequence MRI scan time in clinical practice will be an important direction for future work.
To Rev #3 W1(Limited baselines): As discussed regarding Rev #1-W1, we aim to tackle the issue of insufficient training data. Supervised approaches, however, necessitate large-scale datasets, rendering them impractical for our specific application scenarios. Therefore, we did not include supervised SR methods in our experimental comparisons.
W2(Slow inference): We fully acknowledge that NeRF-based methods are computationally intensive. To address this, CSANeRF has been optimized, significantly reducing the full convergence time to about 1.7 h per volume, compared to 2.5 h of the prior CuNeRF. Inference requires only 3–4 min to generate the final HR volume.
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.
The reviews on this paper are clearly split. The issues brought up by Reviewer #1 overlap with stated weaknesses of the other two reviewers and should be addressed in rebuttal. This rebuttal should focus on 1) a clear description of data splitting and justification of test set independence, 2) description of acquisition parameters (slice profile, voxel size, pulse sequence parameters, etc.), and 3) explain the lack of comparisons to other methods.
- 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 rebuttal provided clear and substantial clarification on both methods and results. All reviewers acknowledge this and recommend acceptance. There are no remaining concerns.
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.
This paper proposes CSANeRF, a cubic structure-aware Neural Radiance Fields method for zero-shot super-resolution of susceptibility-weighted imaging, aiming to address the thick-slice problem and improve the preservation of fine anatomical structures in clinical SWI scans. The reviewers recognized the clinical relevance of the problem, the advantage of the zero-shot setting without paired HR training data, and the technical contribution of combining cube-based sampling with Transformer-enhanced NeRF to better capture spatial correlations. The experimental design was also considered reasonably complete, including reconstruction quality evaluation, ablation studies, and downstream CMB segmentation validation. The reviewers raised concerns about data splitting, acquisition parameters, comparison with GAN/diffusion or supervised SR methods, validation on real LR acquisitions, and inference speed. In the rebuttal, the authors clarified that the method is trained on each subject’s LR volume in a self-supervised zero-shot manner, with HR data withheld for final evaluation. They also justified the comparison with zero-shot NeRF-based and interpolation baselines, provided acquisition details, explained the use of the CuNeRF degradation protocol, and reported the practical inference time. These clarifications effectively addressed the major concerns, and all reviewers supported acceptance after rebuttal. Therefore, I recommend acceptance.
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.
All reviewers recommend to accept this paper post-rebuttal. I agree with this assessment.
The authors submitted a strong rebuttal clarifying the experimental setup / dataset choices and why comparison to supervised SR methods is not applicable in that application.
This is a solid paper. The ablation studies as well as comprehensive evaluation not only on SWI but also on T1w images plus the actual downstream application (CMB lesion segmentation in SWI) convincing.