List of Papers Browse by Subject Areas Author List
Abstract
3D medical image generation is crucial for data augmentation and downstream analysis but remains computationally expensive.
Existing voxel-based generative models operate on fixed-resolution grids, where the latent representation is coupled with output resolution.
This leads to cubic scaling of computational costs and inability to generate volumes at varying resolutions without retraining.
Triplane representations offer a more compact and efficient alternative.
However, current triplane-based approaches either require costly instance-level optimization or are constrained by discrete voxel supervision, preventing resolution-agnostic generation.
To address these challenges, we propose MedTriFlow, an efficient generative framework that bridges a discrete triplane latent space with a continuous anatomical field.
The framework employs a triplane-based autoencoder with 3D-aware projection to compress volumetric data into a compact latent space, where generative modeling is performed for efficient generation.
A continuous anatomical field, modeled by an implicit neural representation, then reconstructs volumetric anatomy at arbitrary resolutions, effectively enabling resolution-agnostic generation beyond the discrete voxel grid.
Extensive experiments on diverse 3D datasets demonstrate that the proposed method, with fast inference (2.6s per volume at $256^3$) and resolution-agnostic generation (demonstrated up to $1024^3$),
achieves superiority over recent 3D generative methods in generation quality, sparse-view CBCT and accelerated MRI reconstruction.
Code and pre-trained models are available at \url{https://github.com/ShanghaiTech-IMPACT/MedTriFlow}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0683_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ShanghaiTech-IMPACT/MedTriFlow
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XuChe_MedTriFlow_MICCAI2026,
author = { Xu, Chenfan AND Dou, Yulong AND Luo, Tao AND Wang, Qian AND Cui, Zhiming},
title = { { MedTriFlow: Efficient Resolution-Agnostic 3D Medical Image Generation with Implicit Triplane Representation } },
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’s main contribution is a novel framework for resolution-agnostic medical image generation. The authors achieve this by introducing a triplane representation of the 3D latent space, which enables efficient and scalable modeling across different image resolutions. Their approach combines a triplane-based autoencoder with a flow-matching U-Net for image sampling. In addition, they incorporate a ControlNet module to enable task-specific and controllable image generation. Through experimentation they showcased that their introduced methodology achieves the best generation performance quantitatively.
- 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 strength are:
1.The paper has a clear structure and is easy to follow. 2.The motivation is clear. Existing generative models are typically limited to a fixed resolution, which is a notable limitation that deserves further attention. 3, The proposed methodology appears reasonable overall. The use of a triplane representation is an interesting way to address the aforementioned problem. In particular, training a triplane-based AE and using a flow-matching model for latent-space sampling is a sensible design choice. 4.The introduction of the triplane AE is a technical novel idea.
- 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 main weaknesses are: 1.The experimentation is a bit limited. The authors only investigate a specific number of CBCT projection views or different MRI acceleration scales. This makes it difficult to determine whether the proposed method consistently outperforms competing approaches under different levels of sparsity. In a future version of the paper, presenting an exploration of different number of views and scales would provide a more complete understanding of its robustness and generalization ability. 2.The proposed methodology appears to contain a larger number of parameters than the other compared methods. That might raise concerns about the fairness of the comparison and the model complexity. A discussion about the trade-off between reconstruction performance and parameter efficiency would strengthen the paper. 3.It is not entirely clear how the task-specific input is projected into the triplane latent representation before being provided to the ControlNet. Adding 1-2 additional sentences explaining this transition would help the reader better follow the overall pipeline.
Minor weaknesses: 1.The pre-trained U-Net is not cited. 2.There are some missing notations in the section 2.
- 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 proposed methodology is sound and generally well presented. However, I have some concerns regarding the experimental setup and certain aspects of the method description. Overall, despite these limitations, I believe the paper makes a valuable contribution and therefore I assign it a weak accept score.
- 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
Review #2
- Please describe the contribution of the paper
The authors propose a resolution-agnostic 3D medical image generation framework, MedTriFlow, which decouples latent representation from volume resolution. The core designs are a triplane-based autoencoder tailored for 3D medical volumes, and an efficient generative modeling in 2D latent space via flow matching. The authors operate in the compressed triplane latent space using flow matching, reducing computational burden and improving inference speed.
- 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.This paper is well-motivated since the limitation of fixed-resolution 3D generation is real and important in 3D medical imaging. 2.The authors integrate triplane autoencoder, continuous anatomical field, and latent flow matching to achieve efficient and flexible 3D generation, which is reasonable. 3.The continuous anatomical field design is interesting, which uses an implicit neural representation to decode triplane features into a continuous 3D field and achieve the resolution-agnostic generation claim.
- 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 “resolution-agnostic” generation claim is somewhat overstated. All training volumes are resampled and padded to 256^3.As a result, the model is not trained on native resolutions. That is, not truly resolution-agnostic learning from native-resolution data. It is unclear whether the 1024^3 generation is anatomically meaningful. 2.Issues about the ablation studies. The authors are suggested to ablation some key designs to make it easier to assess which designs are essential., e.g., flow matching vs. latent diffusion, and the implicit anatomical field vs. a conventional voxel decoder. 3.About how to verify the generation quality, in addition to the generation and reconstruction metrics, more verification can be performed, e.g., Turing tests, use generated images for downstream classification tasks. 4.In terms of native resolution 3D image generation, has anyone tried the video generation approach? The author could discuss the advantages and disadvantages of these two approaches.
- 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.
- 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?
Except for the weaknesses, it is still a well-executed paper that introduces an interesting direction for 3D medical image generation.
- 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’ feedback partialy address my concerns. I keep my rating.
Review #3
- Please describe the contribution of the paper
The paper presents MedTriFlow, a 3D medical image generation framework that (1) performs generation in a compact triplane latent space, and (2) uses an implicit neural representation to decode volumes via coordinate queries, aiming to enable resolution-agnostic output. It also adds structure-guided conditioning for inverse problems such as sparse-view CBCT and accelerated MRI.
- 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.Resolution-agnostic output: INR-based continuous field decoding enables an arbitrary-resolution volume. 2.Flow matching in compact triplane latent achieves fast inference.
- 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.Novelty is limited. Triplane-based compact reconstruction with diffusion models is not new, and the proposed triplane design does not show a clear structural improvement over prior work (e.g., Blaze3DM). The main contribution appears to be mapping a triplane latent to an INR-based field to enable resolution-agnostic reconstruction. However, arbitrary-resolution reconstruction via INR is conceptually close to LIIF, and related LIIF-style ideas have already been studied in medical imaging (e.g., “MedIL: Generating Arbitrary-Resolution Medical Images with Implicit Latent Spaces”). Overall, the authors should more explicitly differentiate the technical designs of both the triplane and the INR components from prior work and clarify why their integration is non-trivial.
2.While higher resolution can be important for diagnosis and surgical settings, it is not clear that arbitrary continuous resolution is necessary. The paper should clarify the specific use case(s) in which a “resolution-agnostic” capability is essential, rather than supporting only a few fixed resolutions.
3.High-resolution benefit is not quantitatively demonstrated. Fig. 3 shows qualitative outputs, but the paper does not report quantitative metrics at higher resolutions. It remains unclear whether coordinate-query high-resolution outputs outperform simple interpolation or super-resolution baselines. Since resolution-agnostic generation is positioned as a key contribution, the paper should demonstrate that its upsampling results are superior to interpolation. If no clear quality gains are observed, the practical value of “resolution-agnostic” capability should be re-justified.
4.Compute cost at high resolution is unclear. Implicit-field decoding typically requires dense per-voxel queries, which can become expensive as resolution increases. Although the paper notes that voxel-based methods are “compounded by the cubic scaling of computational costs with resolution,” the proposed approach may face a similar issue when generating high-resolution outputs. A fair runtime comparison against interpolation baselines at higher resolutions would strengthen the claim.
5.For downstream tasks, the method uses a coarse reconstruction as a structural condition. I am curious how this conditioning enforces consistency between the output and the observed data, and whether any explicit data-consistency loss is required. For example, is the final output explicitly constrained to match the observed data after applying the corresponding degradation operator?
6.The ablation study mainly varies LPIPS and adversarial losses, showing improved perceptual reconstruction quality. However, these loss terms are widely used and do not constitute the paper’s core contribution; they are better framed as training details rather than key mechanisms. The authors should instead design ablations to validate the effectiveness of the proposed modules, especially the components critical to the “resolution-agnostic” mechanism.
- 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.
- 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?
Not applicable
- 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
None
- 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 paper’s core technical contribution appears insufficiently novel. For INR, it provides neither a theoretical justification nor experimental evidence that the INR-based continuous field is necessary or uniquely beneficial for this task.
- 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.
As noted in my previous review comment. The author did not fully justify the weakness of their method.
Author Feedback
We thank the reviewers and respond to the main concerns below. Q1.Meaning of resolution-agnostic generation (Meta-reviewer, R2, R3). Downstream 3D medical tasks live on heterogeneous grids. Dental CBCT, whole-body CT, and MR sequences differ in voxel spacing, resolution, and FOV, and reconstruction grids follow scanner geometry rather than the training grid of a generative prior. A model tied to a fixed grid cannot serve these settings without retraining. Our claim is not native-resolution training, but that one latent code can be decoded at the spacing, resolution, and FOV a downstream task requires through a continuous anatomical field. The encoder still operates on a normalized 256³ grid, which we will state in Sec. 2.Extending encoding to ingest variable native spacing and FOV is our next step. Video-generation-style approaches are also relevant; they treat a volume as an ordered slice sequence, whereas we model it as a 3D implicit neural field from compact triplane features. Q2.Quantitative gain and runtime at high resolution (Meta-reviewer, R3). To verify a real gain at 512³, we compared continuous field decoding with upsampling baselines on both tasks. Each baseline runs at 256³ and is upsampled to 512³ by trilinear or tricubic interpolation, or by a task-specific super-resolution model (Dn-Dp for CT, ArSSR for MRI); ours decodes the same latent directly at 512³. Sparse-view CBCT (PSNR/SSIM), ours 25.94/0.7568, trilinear 23.86/0.7348, tricubic 23.34/0.7335, Dn-Dp 25.07/0.7412.Fast MRI, ours 27.46/0.8651, trilinear 24.97/0.8365, tricubic 24.42/0.8307, ArSSR 26.42/0.8487.Continuous field decoding thus surpasses both interpolation and super-resolution baselines. Latent generation is resolution-independent (~1.2 s); decoding scales with queried voxels (1.4 s at 256³, 4.1 s at 512³, 24.3 s at 1024³), so generation stays in compact 2D latent space. Q3.Novelty relative to Blaze3DM, LIIF, MedIL (Meta-reviewer, R3). Our contribution is the joint design rather than triplane or INR alone. Blaze3DM uses per-volume optimized triplane embeddings and does not learn an encoder-based shared latent space for generative modeling. LIIF and MedIL target image-level continuous representation and native resolution autoencoding without an efficient 3D generative prior. We couple a triplane autoencoder, flow matching in the shared 2D triplane latent, and continuous field decoding within one trainable pipeline. Triplane features and the INR decoder are jointly optimized, so the latent stays compressible for flow matching yet expressive for queries off the training grid. Because the embedding is 2D, the backbone scales with standard 2D diffusion scaling laws and avoids the cubic cost of voxel-based 3D generators. We will clarify this distinction. Q4.Robustness and architectural ablation (Meta-reviewer, R1, R2, R3). 20-view CBCT 26.13/0.7514 and 12× MRI 27.09/0.8623 show the framework remains competitive under heavier undersampling. Replacing our continuous field decoder by a voxel decoder drops reconstruction from 30.78/0.9051 to 29.02/0.8818, isolating the decoding design from the perceptual and adversarial losses already reported. Further comparison with latent diffusion would provide additional insight. Q5.Conditioning, complexity, clarity (Meta-reviewer, R1, R3). The task-specific input passes through the trained triplane encoder and is injected into the flow matching network via the ControlNet branch. We frame downstream reconstruction as structure-guided generative reconstruction; the coarse reconstruction supplies anatomical structure and data fidelity is learned from paired data rather than imposed by a final degradation consistency loss. Although our model has more parameters than several baselines, Table 1 reports the fastest inference and strongest PSNR and SSIM, so the capacity is effectively used. The missing U-Net citation and notation issues in Sec. 2 will be corrected.
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 mixed scores 3, 4, 4.Concerns include the presentation clarity, the technical novelty, the limited experimental evaluations, and potentially overstated claims. AC invites the authors for a rebuttal to carefully address the concerns raised by all reviewers accordingly.
- 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.
This is a borderline paper. R3 insists on rejection and mentioned that the authors’ rebuttal did not address their concern. All reviewers expressed concerns regarding insufficient ablation and unfair comparisons to demonstrate the usefulness and novelty of the proposed method. In the authors’ rebuttal, they agree that their method is a combination of components — which necessitates more thorough comparisons. Overall considerating, AC leans towards rejection, and encourages the authors to add more experiments to justify their method’s effectiveness in future submissions.
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.
Although the technical novelty is limited, the paper is well organized and takes an interesting direction in 3D medical imaging.
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.
While some reservations remain, particularly around the strength of the resolution-agnostic claim and the need for clearer justification of certain design choices, I find the paper’s overall contribution sufficient for acceptance. The work addresses an important limitation in 3D medical image generation, proposes a reasonable triplane/implicit representation framework for efficient and flexible generation, and opens an interesting direction for resolution-agnostic medical image synthesis. I encourage the authors to strengthen the final version by more clearly explaining the practical benefits of the resolution-agnostic setting, better justifying the proposed components, and clarifying the remaining methodological limitations.