List of Papers Browse by Subject Areas Author List
Abstract
Recent diffusion-based approaches have achieved high-fidelity synthesis of missing sequences from partially observed multi-sequence MRI. However, most existing methods rely on a one-way conditional paradigm, in which observed sequences are used as fixed conditions while denoising pre-specified targets, which limits adaptability to arbitrary missing patterns. A more flexible formulation should encode the observed-missing asymmetry within the diffusion process itself. Hence, we propose AsynDiff, an asynchronous-timestep diffusion framework for multi-sequence MRI fusion and synthesis. At each training iteration, we assign sequence-specific timesteps, deliberately inducing inter-sequence noise disparities that promote dependable guidance from low-noise observed sequences to high-noise missing targets. To exploit this asymmetry, we propose cross-sequence noise-aware attention, which adaptively modulates attention weights based on relative noise levels, amplifying reliable features while suppressing noisy and potentially misleading cues. Experiments on BraTS2021 and BraTS-GLI (2024) demonstrate that AsynDiff outperforms comparisons in missing-sequence synthesis. Further analyses across sampling strategies and attention visualizations under varying noise schedules corroborate both the effectiveness and interpretability of the proposed design, highlighting asynchronous-timestep diffusion as a principled mechanism for robust multi-sequence MRI completion. Our code is publicly available.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2167_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/fiy2W/AsynDiff
Link to the Dataset(s)
BraTS2021: https://www.cancerimagingarchive.net/analysis-result/rsna-asnr-miccai-brats-2021/
BraTS-GLI (2024): https://www.synapse.org/Synapse:syn53708249/wiki/626323
BibTex
@InProceedings{HanLuy_AsynDiff_MICCAI2026,
author = { Han, Luyi AND Min, Pengxiang AND Xie, Xinghe AND Zhang, Tianyu AND Wang, Xin AND Gao, Yuan AND Lu, Chunyao AND Liang, Xinglong AND Duan, Yaofei AND Portaluri, Antonio AND Rasoolzadeh, Nika AND van Dijk, Jarek AND He, Muzhen AND Sun, Yue AND Lam, Chan-Tong AND Mann, Ritse AND Tan, Tao},
title = { { AsynDiff: Asynchronous-Timestep Diffusion with Noise-Aware Attention for Multi-sequence MRI Synthesis } },
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
In this manuscript, the authors propose a diffusion strategy for multi-sequence MRI image synthesis based on the JiT architecture. Specifically, unlike traditional diffusion models that treat observed sequences as static conditions embedded to guide the diffusion process, the authors treat observed sequences as variables subject to noise. They introduce an Asynchronous Timestep Diffusion mechanism where observed sequences are perturbed at lower noise levels while missing sequences are perturbed at higher noise levels; the model then predicts the clean output from this complete multi-sequence input. Furthermore, to effectively exploit the disparity in noise levels across modalities, the authors introduce a Cross-Sequence Noise-Aware Attention module within the backbone that dynamically adjusts attention weights based on relative noise levels. Training on public datasets demonstrates that this design achieves superior synthesis performance compared to existing state-of-the-art 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.
1）High Innovation: The authors propose a novel diffusion strategy tailored for multi-sequence MRI synthesis. Unlike existing methods (such as ResViT which relies on one-hot embeddings for fixed sequence conditioning), this work introduces an asynchronous noise mechanism that allows observed sequences to participate in generation in a low-noise, non-fixed state. Combined with a cross-sequence attention mechanism that dynamically learns relative noise levels across modalities, this approach enhances the robustness of the missing sequence synthesis task compared to traditional fixed-condition inputs. 2）Strong Scalability: The core mechanism of asynchronous noise proposed by the authors is model-agnostic. Its core principles are not dependent on the specific diffusion backbone used and can theoretically be adapted to various diffusion architectures. 3）Thorough Analysis and Focus: The authors provide a detailed discussion of ablation studies in the experimental section, elucidating the rationale behind the architectural design choices. Additionally, the proposed method is comprehensively compared against several state-of-the-art methods with different architectures on the same task. The experimental results are abundant and well-documented.
- 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) Concerns regarding Ablation Study Design: In Table 1, the w/o Asyn option appears, based on the text, to retain the proposed Cross-Sequence Attention module rather than serving as a “native JiT backbone” where all introduced modules are removed. Given that JiT is a high-performance denoising architecture, it could potentially perform competitively on this task even with its original synchronous timestep and condition embedding methods. The lack of an experiment comparing against a “Pure JiT (Synchronous)” baseline makes it difficult to disentangle whether the performance gain is driven by the backbone itself or the proposed asynchronous mechanism. This raises the question: is the JiT backbone alone sufficiently powerful to surpass the baselines compared in the paper, or is the contribution of the proposed concept marginal? 2) Lack of Discussion on Model Complexity and Parameters: As an independent noise sampling strategy not tied to the diffusion architecture, the authors’ asynchronous noise mechanism and cross-sequence attention module must work in synergy. However, the paper lacks a quantitative analysis of the additional parameters and computational complexity introduced by these two mechanisms. As seen in Figure 2, introducing a cross-sequence attention block after every attention module implies a significant increase in parameters and computational overhead. This potential cost is not sufficiently discussed, particularly for resource-constrained application scenarios. 3) Reproducibility Concerns: The authors claim to make source code publicly available, but as of now, no link has been provided, preventing verification of the code’s reproducibility.
- 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?
The proposed asynchronous noise mechanism and cross-sequence attention module based on the diffusion architecture demonstrate significant innovation and scalability. The manuscript is well-structured with sufficient experimental evidence. Although there are shortcomings in the rigor of the ablation study and the assessment of model resource consumption, the core concepts are worthy of further in-depth research.
- 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
This paper proposes AsynDiff, an asynchronous-timestep diffusion framework for missing-sequence multi-sequence MRI synthesis. Instead of using the same noise timestep for all sequences, the method assigns sequence-specific timesteps to create noise asymmetry, so that less corrupted observed sequences can better guide the reconstruction of missing targets. The paper also introduces a cross-sequence noise-aware attention module and shows improved results on both BraTS2021 and BraTS-GLI.
- 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 addresses a practical and important problem in multi-sequence MRI, where missing modalities are common in real clinical settings. The main idea is clear and well motivated: asynchronous noise levels provide a simple but effective way to encode observed-missing asymmetry into the diffusion process. The proposed noise-aware attention is also intuitive and improves the interpretability of cross-sequence information transfer. Experimentally, the paper includes comparisons with several GAN- and diffusion-based baselines, ablations, inference analyses, and external validation on BraTS-GLI, which makes the empirical section reasonably strong.
- 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.
Although the experiments are solid, the gains over the strongest diffusion baselines are generally not very large, so the practical impact may be somewhat incremental (especially the robustness of the model). The evaluation is also limited to BraTS-style brain MRI benchmarks, and it remains unclear how well the method would generalize to other anatomies, scanners, or broader multimodal synthesis settings.
- 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?
This manuscript presents a clear and technically sound method for an important MRI synthesis problem, and the experimental results are consistent and reasonably comprehensive. The asynchronous-timestep design is simple but well motivated, and the inclusion of external validation and interpretability analysis strengthens the paper. However, the overall novelty is moderate and the empirical gains over strong recent baselines are not dramatic. Overall, I find the paper marginally above the acceptance.
- 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 #3
- Please describe the contribution of the paper
This paper proposes AsynDiff, an asynchronous-timestep diffusion framework for multi-sequence MRI synthesis. The core contribution lies in breaking the conventional “synchronous noising / fixed condition” constraint in multi-modal diffusion models. By assigning independent noise timesteps to different modalities during training and introducing a “Cross-Sequence Noise-Aware Attention” mechanism, the model adaptively leverages reliable features from relatively low-noise sequences to guide the generation of high-noise target sequences.
- 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 design of asymmetric timestep noising is highly inspiring. Embedding the asymmetry of “observed vs. missing” data directly into the diffusion dynamics—rather than merely treating it as static condition channels—represents a substantial breakthrough over the existing one-way conditional generation paradigms in medical imaging. 2) The design of the Cross-Sequence Noise-Aware Attention is elegant. By utilizing the global timestep tall as the modulation parameter for AdaLN, the network precisely regulates the information flow in the latent space. 3) The attention heatmaps in Figure 6 are intuitive and compelling. They clearly demonstrate that information flows from low-noise sequences to high-noise sequences as expected.
- 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) Relying solely on standard computer vision metrics (PSNR, SSIM, LPIPS) is insufficient. Synthesized MRI sequences must be proven clinically reliable, particularly concerning lesion boundary fidelity and artifact avoidance. The authors completely omit utility evaluations on downstream clinical tasks. If the synthesized images induce hallucinations or degrade downstream segmentation accuracy, the clinical translational value of this method is severely undermined. 2) In Table 2, the performance of the 3D models is noticeably inferior to their 2D counterparts. The severely restricted global receptive field caused by the small 8×8×8 patch size leads to local contrast bias. 3) The experiments are strictly confined to the BraTS 2021 and BraTS-GLI 2024 datasets. Without validation across different clinical centers or different pathologies, it is difficult to ascertain whether the model has learned universal cross-sequence physical mappings or merely overfitted to the BraTS glioma distribution. 4) While Figure 4 compares four inference sampling strategies, the main text only provides vague linguistic descriptions. The complete absence of strict mathematical formulations defining the tm trajectories within the Euler ODE solver severely compromises the methodological rigor and reproducibility.
- 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.
- 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 decision for a “Weak Accept” is primarily driven by the profound depth and high novelty of the generative mechanism. The methodology aligns beautifully with physical intuition. However, its evaluation protocol is overly skewed towards traditional computer vision standards, alongside a somewhat suboptimal 3D extension. If the authors can provide the downstream tumor segmentation evaluation in the rebuttal—proving that the synthesis does not induce medical artifacts or information distortion—I would be very glad to raise my score to an “Accept”. Conversely, failure to provide this will warrant maintaining the score or opting for rejection.
- 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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 reviewers find AsynDiff conceptually elegant and technically solid: asynchronous timesteps and noise-aware cross-sequence attention offer a fresh, intuitive way to model observed–missing modality asymmetry, with consistent gains.
There is some hesitation about depth of validation: improvements over strong baselines are modest, evaluation is confined to BraTS-style settings with standard image metrics (no downstream or clinical utility), ablations and complexity analysis are incomplete.