List of Papers Browse by Subject Areas Author List
Abstract
Clinical cervical Magnetic Resonance Imaging (MRI) routinely relies on multi-view anisotropic 2D scans with large slice gaps, causing severe through-plane information loss that hinders 3D diagnosis. Reconstructing an isotropic 3D volume from such scans is challenging due to heterogeneous intensity distributions and partially mismatched fields of view (FOVs). We propose an anatomy-style decoupled hybrid implicit neural representation (INR) framework for multi-view super-resolution without paired high-resolution supervision. Our approach aligns orthogonal scans within a unified coordinate system, utilizes a two-stage feature extraction module to inject anatomical priors, and introduces a stochastic view dropout strategy with learnable style embeddings to explicitly disentangle shared anatomy from view-specific intensities. Extensive experiments on 120 clinical cervical MRI pairs demonstrate artifact-free isotropic 3D reconstruction with 50% rated as excellent by a blinded clinical expert, achieving axial and sagittal FIDs of 45.14 and 55.50 and a Lesion Dice of 0.673. Our code is available: https://github.com/zhibaishouheilab/JointSR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1142_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zhibaishouheilab/JointSR
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaQi_Reconstructing_MICCAI2026,
author = { Zhang, Qi AND Chen, Xiuyuan AND Hu, Yibo AND Shen, Hongxing AND Sun, Jianqi},
title = { { Reconstructing Isotropic 3D Cervical MRI from Anisotropic Clinical Scans via Anatomy-Style Decoupled INRs } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a novel decoupled hybrid INR framework for generating an isotropic, high-resolution volume from unaligned, anisotropic input scans, without reliance on paired HR data. The authors provide a strong evaluation with a good mix of baselines and clearly discuss the limitations.
- 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 is well-written and provides an intuitive understanding of the proposed method 2.The paper proposes a unified coordinate system to accommodate various input views (scans) for the subsequent INR to learn a good-quality mapping from input coordinates to signal values. While it may seem straightforward, I appreciate the novelty of this specific module as it directly impacts information coverage captured by the INR. 3.The stochastic view doprout is a quite interesting strategy to elegantly handle single-view FOV conditions during inference and will make the entire pipeline more robust. 4.I also appreciate that the method works with sparse-input points, unlike other INR approaches that tend to use all coordinates as inputs, showing that the proposed method is robust to sampling and generalizes well. 6.Authors provide a clinical evaluation of their proposed method demonstrating high-quality, clinically aligned impact. 5.Limitations of the work are also well discussed and acknowledged.
- 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.Cover figure (Figure 2) could’ve been a bit clearer. Currently, the figure has a lot of information flows (and arrows), making it visually cluttered. 2.Table 1 (axial view column): It seems that the proposed method scores poorly. Can the authors explain why cubic performs so much better than the proposed method? 3.Furthermore, can the authors clarify the ablation results in Table 1 (axial view) and explain why ablations reflect better performance compared ot the proposed method? 4.Is the style embedding space interpolatable and potentially convex? Do the authors have any insight in the interpretability of the style embedding?
- 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.
(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 highlights a strong new method of generating HR isotropic volumes from anisotropic input scans. The paper also shows clinical value. The results in Table 1, however, raise some questions and require further clarification.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes a novel approach to reconstructing unified arbitrary-resolution representations from multiple anisotropic views. Unlike previous work, the paper formulates their approach under a multi-subject setting, allowing for view-dependent intensities to be handled via learnable style embeddings.
- 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 extends previous work by formulating their work under a cohort-based training scheme. -The methodology is intuitive and straight-forward. The architecture does not appear to be out of reach for general clinical settings in terms of computational demands. -The paper offers strong qualitative visualization, showcasing the visual fidelity of their method, as well as segmentation maps derived from their interpolated predictions. -Ablation is performed on the core design choices behind their 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.
While ablation is provided, both quantitatively and qualitatively, little theoretical discussion is present on why an initial single-view super-resolution is beneficial. The authors mention this approach offers reduced blurring effects to traditional interpolation, but why is this necessary in the first place? Why can’t encoder features be directly extracted from the raw anisotropic scans? This initial interpolation feels somewhat to work against the spirit of the paper, which is to create an improved interpolation by informing the process with multiple complementary views.
- 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
Is there a specific reason why trilinear interpolation is used for feature interpolation? The commonly termed ‘curse of dimensionality’ guarantees that interpolation in higher dimensionality spaces becomes less trivial specially as feature grids become coarser. The effect of this may be negligible in the author’s setting, but a discussion on this issue could be appreciated for ablation in future work regarding non-linear (or even learnable) interpolation. If the need for the initial intensity interpolation is to decrease the anisotropicity of the resulting feature grid, perhaps this may contribute to a solution.
Perhaps a further discussion on the reason MR intensities vary depending on slice orientation (MR physics, scanners obtaining images in frequency space, etc) would prove useful for readers, as not every imaging modality suffers from this issue.
Papers from eg. fetal MRI [1] try to explicitly include some of these MR-specific physical properties into forward propagation of the model when learning the superresolved representation. These include slice-wise bias fields, modeling voxels as point-spread functions, etc. These MR-specific imaging properties might be interesting to include for interested readers in the camera ready version. Nonetheless, in my opinion, the author’s approach to model these properties implicitly as fully data-driven learnable conditioning embeddings is very interesting and efficient.
[1] - Xu, Junshen, et al. “NeSVoR: implicit neural representation for slice-to-volume reconstruction in MRI.” IEEE transactions on medical imaging 42.6 (2023): 1707-1719.
- 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. Methodology is intuitive, simple, yet novel. Results are well-presented and validate the hypothesis described by the authors. Visualizations are neat and offer useful information on the quality of the results. Ablation is thorough.
- 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 author’s provided clear arguments on their rebuttal. My methodological questions were answered.
Moreover, I find it comforting to see that the author’s are knowledgeable enough on the general MR imaging field to concisely address reviewer #4’s concern regarding a lack of ground-truth super-resolution volumes. Obtaining a ground-truth high-resolution volumes isn’t trivial. The slice-wise nature of MRI acquisitions creates partial volume effects that are orientation dependent. Acquiring more intermediate slices will not produce a meaningful GT you can approximate from an orthogonal acquisition. Furthermore, meaningfully increasing in-plane resolution requires using a separate scanner altogether, which additionally add non-trivial overhead volume registration issues.
Lastly, the author’s acknowledged my feedback on physics informed literature and offered to further discuss these details in the camera-ready version.
I find the authors’ rebuttal satisfactory.
Review #3
- Please describe the contribution of the paper
Obtaining an isotropic high resolution MRI scan from stitching multiple FoV MRI from different (lower resolution) inplane acquisitions is difficult, so the authors present an implicit neural representation (INR) framework to do this accurately. It appears that the framework involves using a patch based approach to sample using unified R^3 coordinates to reassemble a synthesized isotropic scan using style-based codes to disentangle the anatomy from intensity, while using a dropout strategy in the overlapping regions.
- 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 use of style and Fourier codes in MRI seems novel, as well as the integration of patch based DRN for features for the INR all without needed paired high resolution training data, which is very rare. The dataset is of cervical MRI (not common) and the model appears to run on a 4090 RTX. Visualizations of figure 3 are nice, as is the clinical expert evaluation of the results of the model. The results are promising showing improvements on one of the views in comparison with other INR methods.
- 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 dataset doesn’t have high resolution images for quantitative measures from ground truth of the super resolution tasks (like PSNR, SSIM etc.) to judge how faithful is the super-resolution, which is critical to avoid artefacts and hallucination. The experiments show that the cubic interpolation out performs all the methods on one of the views for some reason. SCL-INR is missing from the clinical evaluation. The proposed method is only compared against two other published methods, not sure why ref 7 was not also compared against. The overuse of sub-headings end up using valuable vertical space and length of the paper, resulting in a very short methods section that could have benefited from more details on model.
- 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 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?
“We collected 120 paired dual-view (sagittal and axial) anisotropic cervical MRI scans from an in-house cohort (randomly split into 100 for training and 20 for testing)”, does not mention ethics in obtaining images etc.
- 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?
The authors lack high resolution ground truth to show quantitative benefits of the method. Even a simple super-resolution task one a benchmark dataset would have benefited the paper as FID and KID only assess fidelity not faithfulness. The comparison could have had one more method. The methods section could have benefited from more details and other details such as how coordinates are important in the model and additional visualizations of the benefits of the proposed model.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and Meta-Reviewer for their constructive feedback. We address concerns by theme.
1.Why Cubic Appears Better on Axial (R2 Q2–Q3, R4 Q2, Meta-Reviewer) Cubic achieves lower axial FID/KID than our method, as expected. Per §3.3, Cubic and SA-INR post-process by histogram-matching sagittal SR to the axial reference and appending non-overlapping regions. Cubic’s axial output is thus the native axial in-plane data, yielding artificially low scores against the same reference (KID=0.000, FID=1.641). This single-view optimization collapses sagittal reconstruction (FID=100.96 vs Ours 55.50; Spine Dice=0.723 vs Ours 0.885). The same mechanism explains why ablations appear better on axial: they overfit the axial distribution while degrading sagittal quality (Dice 0.558 w/o SVD vs Ours 0.885). Our method is explicitly designed for balanced bi-plane reconstruction. Lesion Dice: Ours 0.673 vs Cubic 0.529, SA-INR 0.532, SCL-INR 0.337, critical for clinical utility.
2.HR Ground Truth, Faithfulness, Artifacts (R4 Q1, Meta-Reviewer) We acknowledge the absence of paired 3D HR GT, a field-wide challenge in clinical MRI super-resolution. Isotropic 3D scans are rarely acquired in routine practice due to long scan times, and no public cervical 3D HR dataset exists. To validate both image fidelity and anatomical faithfulness, we have employed a triple-validation strategy: (1) FID/KID against native in-plane slices for distribution-level fidelity; (2) blinded clinical expert evaluation on all 20 test cases using a 5-point Likert scale (5 = excellent, artifact-free; i.e., free of hallucination, stitching, and intensity artifacts), with our method achieving 50% scored 5/5 and all ≥3, the expert specifically assessed for clinically meaningful artifacts; (3) downstream segmentation Dice as an anatomical faithfulness proxy (Lesion Dice 0.673 vs Cubic 0.529, SA-INR 0.532). These three complementary strategies collectively substitute for the missing PSNR/SSIM and directly address concerns on artifacts and faithfulness. Future work includes recruiting subjects for isotropic 3D MRI scans to enable direct PSNR/SSIM computation.
3.Method Design and Baseline Comparisons (R3 Q1, R4 Q3–Q5, Meta-Reviewer) Single-View SR Prior (R3): Directly learning 3D features from raw LR scans with large slice gaps is inherently challenging and our single-view SR provides a “warm-up”. Unlike simple interpolation, it recovers anatomically plausible structures even when inputs contain noise or artifacts, yielding gains (Lesion Dice: 0.634→0.673; Spine Dice: 0.841→0.885). The SR priors serve only as feature extractor inputs.
Baseline Exclusions (R4): SCL-INR was excluded from clinical evaluation because quality was too poor for meaningful assessment (Dice=0.513 vs Ours 0.885; Fig. 3). Ref [7] was excluded because it requires identical input dimensions, incompatible with variable-FOV cervical data.
Method Details (R4): We will reduce sub-headings, expanding the unified coordinate system and feature extraction pipeline.
4.Limitations, Presentation & Ethics (R2 Q1,Q4; R3 Q2–Q3; R4 Q6) Trilinear Interpolation (R3): At current resolution, trilinear interpolation suffices; quality may degrade at coarser grids (curse of dimensionality). We will add discussion of this limitation and identified learnable interpolation as future work.
MR Physics & NeSVoR (R3): We will add discussion of MR intensity variation across slice orientations (frequency-encoding, coil sensitivity) and cite NeSVoR [Xu et al., IEEE TMI 2023] as complementary work.
Style Embedding (R2): The MLP decoder is continuous; linear interpolation s(alpha)=(1-alpha) s_A+alphas_B therefore guarantees continuous intensity transitions, which we have verified.
Figure 2 (R2): we will simplify it to reduce visual clutter.
Ethics (R4): IRB approval and informed consent were obtained; inadvertently omitted, now added.
We trust these clarifications and revisions address all concerns.
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 received multiple favorable reviews. Weaknesses were noted and were consistent between reviewers. Concerns about fidelity are founded but alleviated by both a rater study and segmentation results. This rebuttal should focus on explaining the positive performance of cubic interpolation in Table 1, clarifying method details, and expanding on the use of the rater study to further alleviate concerns about artifacts.
- 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 provided a substantial and convincing rebuttal that answered each concern well. They clearly understood the reviewers’ comments and addressed each one. I believe that all reviewers would recommend acceptance at this stage. I commend the authors on their detail in addressing the comments.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The rebuttal includes new quantitative results and additional baseline comparisons that were not part of the submitted manuscript, which violates the rebuttal policy and warrants desk rejection.
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.
This is a good-quality submission handling anisotropic clinical datasets. The authors did a good job in clarifying details. I vote for accept considering the consistent positive comments.