List of Papers Browse by Subject Areas Author List
Abstract
Magnetic resonance imaging (MRI) is vital for diagnosing abdominal and neurological conditions, yet conventional sequential slice acquisitions favor in-plane over through-plane resolution to minimize scan time and motion artifacts, leading to anisotropic data and reduced volumetric accuracy. Existing super-resolution (SR) techniques reconstruct isotropic images from anisotropic scans but often rely on simulated downsampling or limited 3D isotropic data, emphasizing through-plane interpolation rather than preserving full anatomy. We introduce SIMPLE, a Simultaneous Multi-Plane Self-Supervised Learning approach that directly restores isotropic MRI from real-world multi-plane acquisitions via adversarial training. Testing on OASIS-1 brain (n = 416) and Crohn’s disease abdominal (n = 115) MRI datasets demonstrates SIMPLE’s superiority in image fidelity and anatomical detail over state-of-the-art methods. Notably, SIMPLE achieved lower averaged Kernel Inception Distance (KID) scores than SMORE4 in both brain MRI (28.709 vs. 29.295) and abdominal MRI (17.435 vs. 20.724), retained higher-frequency details as confirmed by Fourier analysis, and was rated 1.5 points higher in the axial plane by radiologists. By improving volumetric analysis and 3D reconstructions, SIMPLE shows promise for enhancing diagnostic accuracy in pathologies demanding precise structural visualization.
Our source code is publicly available at https://github.com/TechnionComputationalMRILab/SIMPLE.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2075_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/TechnionComputationalMRILab/SIMPLE
Link to the Dataset(s)
OASIS-1 dataset: https://sites.wustl.edu/oasisbrains/home/oasis-1/
BibTex
@InProceedings{BenRot_SIMPLE_MICCAI2025,
author = { Benisty, Rotem and Shteynman, Yevgenia and Porat, Moshe and Ilivitzki, Anat and Freiman, Moti},
title = { { SIMPLE: Simultaneous Multi-plane Self-supervised Learning for Isotropic MRI Restoration from Anisotropic Data } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15972},
month = {September},
page = {551 -- 561}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduced a adversarial framework to reconstruct isotropic MRI volumes from clinical anisotropic multi-plane acquisitions. It contains 3D UNet generator and 3 2D Discriminator that incorporate all views (or two of them) input with the 3D generated volume for 3 L1 losses.
- 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 method avoids synthetic downsampling or fully supervised training by using a 2D SR pretrained model to generate plane-specific pseudo HR references for GAN.
- Results showed improvement over SMORE4, linear interp baseline, ATME in KID, FID, Likert Scale, etc.
- Good motivation on the multiview setup that requires no isotropic ground truth.
- 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 biggest issue is also part of the strength, which is lack of ground truth. The authors made a lot of effort on this but apparently were hand-tied by their private dataset, which make the contribution very limited to the general audience. There were many good supervised volumetric throughplane SR works in the recent several years in medical imaging and these methods can be easily adapt to MRI, like SAINT, TVSRN, ResVoxNet, CuNeRF, CycleINR, etc. SMORE is a rather old benchmark. At least the authors should try to compare these using OASIS data with downsampled LR input. The evaluation is also a problem given lack of ground truth, depending on perceptual metrics, Fourier transforms, radiologist scores, which are helpful but subjective or indirect. Lack of voxel level validation reduces its clinical impact. Given the limitation in objective validation, the paper lacks of downstream task as a surrogate evaluation to prove the efficacy.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
ill-defined training and evaluation which limits the method’s generalizability. lack of comparison with recent works.
- Reviewer confidence
Very confident (4)
- [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.
I still this method although valid, is built on top of a lot assumptions, i.e. trying to tackle a hard problem even though there are better approaches with better settings (using large real paired HR LR data like TVSRN). The authors brought up concern about hard to obtain real data in clinical setting and scan time, cost etc. But that’s the whole reason why people develop methods to apply in these real world settings. the authors don’t have to worry about these because you are just inferencing in real world, not training.
Review #2
- Please describe the contribution of the paper
The paper introduced SIMPLE, a Simultaneous Multi-Plane Self-Supervised Learning approach that directly restores isotropic MRI from real-world multi-plane acquisitions via adversarial training. It shows better results compared to previous multi-plane super resolution methods.
- 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 paper combines a 3D generator with several 2D discriminators in its GAN training, which is a novel approach. The proposed method can better utilize the high-resolution 2D images for this 3D super resolution network.
-
The proposed method can obtaion good results in different plane with a single model.
-
- 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.
-
In Figure 1, the description is not clear. Such as “Dashed arrows denote sampling along each plane”, there are many dashed arrows, the authors should specify the color of arrows.
-
The network uses a pretrained 2D super resolution model to generate high resolution images, and treat them as “real” high resolution slices. This approach may propagate errors. The author should discuss more on the pretrained 2D model.
-
In Figure 2, the captions over the images are very confusing and misleading. The “Axial Volume”, “Coronal Volume” actually mean the images are interpolated from those volumes, but readers may think they are displaying those volumes. Those captions need to be revised.
-
The paper should show a table for the Likert scale results.
-
I don’t understand why this method is a self-supervised method. I think the pretrained 2D super resolution model is trained with high resolution labels. Then the proposed method should be a supervised learning method to some degree.
-
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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 — could be 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 has shown some novelty and better results than previous methods. But its overall presentation is not clear enough.
- 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 have addressed my concerns in the rebuttal.
Review #3
- Please describe the contribution of the paper
MRI scans are generally acquired in anistropic resolution in practice with a loss of information in the planes in the plan not used for acquisition which gives suboptimal 3D reconstruction. The main contribution in this paper is a self-supervised method called SIMPLE that is able to reconstruct isotropic MRI volumes from anisotropic volumes to improve diagnostic accuracy without relying on simulated downsampling.
- 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 main strengths of this paper are:
1) Direct isotropic reconstruction with multi-plane self-supervision: this framework ensures consistency in the three plances while avoiding synthetic downsampling and single-plane interpolation.
2) Adversarial Learning: this framework includes several 2D discriminators for each plane using a GAN approach. This ensures the reconstructed slices preserve the specific features and high-frequency details proper to the anatomical structure.
3) Validation on two datasets: the study is evaluated over two distinct datasets : OASIS-1 Dataset (brain) and Abdomen Dataset showing its adaptability to different body structures.
- 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) Computational cost: the model involves 3D patch extraction, multiple discriminators and several losses. This could lead to high memory consumption and significant computational requirements compared to more traditional methods with single-plane interpolation. 2) Ablation of the different components: there is no assessment of the contribution of each of the components : discriminators, ATME-generated references and hyperparameters.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
Given that the model involves extracting 3D patches and multiple discriminators, this potentatially means greater computational ressources. A cost analysis comparing the current model and traditional methods relying on simple interpolation is missing.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This work treats a relevant challenge in the field of medical imaging for restoring isotropic MRI from anisotropic MRI at time of acquisition. The framework using multi-plane self-supervision and avdersarial leanring ensures preservation of high-frequency details. Moreover, the validation on two different datasets shows the robustness of the current method. However, in terms of reproducibility, a study on the cost associated with these components would be needed.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 for their insightful feedback. We appreciate the positive remarks on the novelty of our approach, which ensures three-plane consistency using clinically available high-resolution 2D images without synthetic down-sampling or supervised training (R1, R2, R3). We also value the recognition of our paper’s clarity (R1, R3) and thorough validation across diverse datasets and methods (R1, R3).
Ground Truth Limitations, Evaluation, and Generalizability (R1): We agree that the lack of isotropic ground truth is a challenge—but it defines the setting SIMPLE is designed to address. Fully isotropic volumes are rare in clinical abdominal MRI due to acquisition constraints; even OASIS (1×1×1.25 mm) is not truly isotropic. SIMPLE is fully self-supervised, needing no isotropic labels or synthetic down-sampling, and generalizes to real-world anisotropic data. In contrast, most cited methods (e.g., CycleINR, SAINT) are supervised and require isotropic ground truth, which involves time-intensive acquisitions not feasible in routine practice—especially in abdominal imaging, where anisotropy is inherent due to scan time, motion, and patient comfort. These methods are not applicable to our setting, which reflects real-world conditions without isotropic references. SIMPLE is designed for this clinical scenario, making such comparisons inappropriate. We clarify that we used SMORE4 (2023), a substantial improvement over the 2020 version. We focused on SIMPLE’s formulation and real-world utility. Results on simulated isotropic data (e.g., OASIS) were omitted due to space, but we plan to include them in an extended version. Due to the lack of voxel-level ground truth, we used qualitative metrics, Fourier analysis, and expert radiologist scoring, which we consider most clinically meaningful. Additional radiologist evaluations are in progress. SIMPLE’s three-plane consistency also acts as implicit regularization, improving robustness without isotropic supervision.
Clinical Relevance (R1): Since voxel-level validation is limited, we emphasized radiologist scoring, with additional readers underway. SIMPLE’s multi-plane structure enhances robustness. We agree that downstream tasks offer a useful surrogate and plan to explore this in future work.
Self-Supervision & Use of ATME (R2): SIMPLE uses real anisotropic slices without synthetic or isotropic labels. ATME is trained per plane to super-resolve through-plane structure using in-plane information only. These outputs serve as soft guidance in SIMPLE, not fixed labels, and are constrained by cross-plane consistency. Since planes are not co-registered, training avoids interpolation artifacts.
Error Propagation from ATME (R2): While generative models can introduce artifacts, each ATME model is independently trained on abundant real HR slices (~45× more slices than cases in abdomen and ~128× for brain). SIMPLE’s three-plane constraint reduces the effect of potential plane-specific errors.
Computational Cost (R3): SIMPLE remains efficient on a 32GB GPU despite 3D patches and multi-discriminator training:
- SIMPLE: 118.3M params, 8046MB peak memory (batch of 16), 7.91s/case
- ATME: 38.6M params, 14,707MB peak (single case), 19.08s/case Patch extraction helps manage memory and patch size is adjustable; discriminators are lightweight and optimized for high-frequency content.
Ablation Studies & Tables (R2, R3): We ablated each discriminator. Axial contributed more (via KID), likely due to higher coronal resolution. The full model performed best. Hyperparameters (α, β, γ, λ) were tuned via grid search. Likert-scale results were omitted due to space but will be added in the final version. Figure 2 compares ATME outputs across methods and planes.
Presentation & Code Availability (R2, R3): We will revise figures for clarity as suggested. Code will be released upon acceptance.
We thank the reviewers again and will incorporate all feedback to further improve the clarity and impact of our work.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
KID/FID/IS are hard to interpret without corresponding ground truth or task-based evaluation, and results of Likert scale are not much discussed. visually, the output images appear quite similar, and evaluating a downstream task would better highlight the strengths of the proposed method.
Meta-review #3
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The authors propose a novel method for using 2D super-resolution networks to super-resolve multi-view LR volumes. Several concerns were brought up during the review. Primarily, there is a lack of quantitative evaluation. The authors contend this is due to a lack of available data, but an isotropic volume can easily be degraded (as in this work) to multiple “2D” acquisitions in different views, and the final result can be compared voxelwise. There is also a lack of quantitative downstream analysis. The rebuttal did not adequately address these concerns. This along with comprehensive ablation studies, will help demonstrate the validity of the method.