List of Papers Browse by Subject Areas Author List
Abstract
Functional magnetic resonance imaging (fMRI) denoising is a crucial preprocessing step in neuroimaging studies, as noise degrades the reliability of downstream analyses. Previous approaches for fMRI denoising either rely on predefined noise patterns or train dataset-specific models, restricting their reliability across various datasets due to inter-dataset variations in scanner types, scanning protocols, and preprocessing pipelines. Additionally, applying previous approaches to new datasets requires extensive expert signal/noise annotations. To mitigate this reliance, leveraging existing datasets to train sparsely labeled datasets is a practical solution, but inconsistencies in labeling criteria hinder effective adaptation. To address these challenges, we propose a meta-learning-based semi-supervised domain adaptation framework, enabling the learning of dataset-irrelevant features from sparsely labeled datasets by leveraging existing labeled datasets with two key components: (1) a dataset-irrelevant feature extractor trained by meta-learning to capture noise patterns across multiple datasets, and (2) dataset-specific classifiers optimized by decoupled training to handle inconsistencies in labeling criteria. Our proposed approach shows outstanding performance on four fMRI datasets in both fully labeled and sparsely labeled conditions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3564_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/KeunsooHeo/metaclean
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HeoKeu_Sparsely_MICCAI2025,
author = { Heo, Keun-Soo AND Han, Ji-Wung AND Bak, Soyeon AND Lim, Minjoo AND Kang, Bogyeong AND Park, Sang-Jun AND Lin, Weili AND Zhang, Han AND Shen, Dinggang AND Kam, Tae-Eui},
title = { { Sparsely Labeled fMRI Data Denoising with Meta-Learning-Based Semi-Supervised Domain Adaptation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15966},
month = {September},
page = {583 -- 593}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a meta-learning-based semi-supervised domain adaptation framework for denoising functional MRI (fMRI) data, aiming to reduce reliance on fully labeled datasets. The key contributions include:
- A dataset-irrelevant feature extractor trained using Model-Agnostic Meta-Learning (MAML) to capture robust noise features across domains.
- A decoupled training strategy for dataset-specific classifiers to handle label inconsistencies across datasets.
- Validation on four diverse fMRI datasets, under both fully labeled and sparsely labeled (10%) settings, showing superior performance compared to a recent baseline.
- 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 is well-written and easy to follow. The decoupled training strategy is well-motivated and aligns with the problem of annotation heterogeneity in neuroimaging studies.
- Performance is compared under both 10% and 100% label settings, demonstrating robustness even in low-label regimes.
- 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 the benefits of meta-learning and decoupled training are separately evaluated, the impact of the alignment loss (L_align) is not isolated.
- The paper evaluates domain adaptation performance but only with fixed datasets. It’s unclear how the method would perform in completely unseen target domains not included during training.
- The feature extractor is said to be based on [18], but no further architectural details are given (e.g., layer numbers, filter sizes, activation functions).
- 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.
- 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?
Although the experimental results are impressive but the technical contributions are not enough for MICCAI standard. This paper can be considered as the ensemble of existing methods (meta-learning, dataset-specific classifier).
- 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 presents a meta learning-based semi supervised domain adaptation framework for fMRI denoising. They focus particularly on learning the dataset agnostic features for improved generalization. Their method learns from a sparsely labeled dataset and they use dataset specific classifiers to capture changes in noise labeling criteria.
- 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 authors present a three part framework that helps them disentangle features from the source and target domain
- The authors evaluate their framework on four different fMRI datasets
- The authors conduct statistical tests to ascertain the significance of their 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.
- The rationale behind the fMRI denoising problem across different domains is clear, however, the rationale for meta-learning and semi-supervised learning is completely missing. Why did the authors choose this specific paradigm.
- The authors compare their method with just one state-of-the-art approach.
- it is unclear why in the first pertaining step the target domain labels were also used.
- 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.
- 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.
(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?
The paper presents a meta learning approach for fMRI denoising but they don’t provide a rationale towards developing their methodology. This makes it hard to position the method in context of other meta learning approaches.
- 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 #3
- Please describe the contribution of the paper
The authors present a meta-learning-based semi-supervised domain adaptation framework for fMRI denoising, designed to work even when only a small portion of the target data is labeled. Their approach introduces two key ideas: first, a dataset-irrelevant feature extractor trained using MAML-style meta-learning to learn noise patterns that generalize across datasets; and second, a decoupled training strategy for dataset-specific classifiers, which helps handle differences in labeling standards between datasets. They test their method on four varied fMRI datasets (HCP, BCP, WHII-MB6, and WHII-STD) and demonstrate clear improvements over a strong CNN baseline, both when using only 10% of the labeled data and in fully supervised settings.
- 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.
This paper implements an important challenge in fMRI ICA denoising—how to work with sparse labels across datasets that often differ in labeling standards. The combination of meta-learning and decoupled training is both novel and well-justified, offering a fresh approach to a long-standing problem. The experiments are thoughtfully designed, using a diverse set of fMRI datasets collected with different scanners, populations, and protocols. Results are presented under both low-label (10%) and full-label (100%) scenarios, with detailed ablations and statistically sound evaluations. The visualizations, particularly Fig. 2 with Grad-CAM, compellingly illustrate how the model focuses on meaningful brain regions. Overall, the paper is clearly written, well-structured, and provides enough detail to support reproducibility.
- 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 proposed approach is well-motivated and demonstrates strong results, but the evaluation could be more compelling with clearer methodological context. For instance, while wavelet-transformed ICs are briefly mentioned in the visual analysis, their role in the overall pipeline isn’t clearly described in the methods section, making it difficult to assess their contribution. Additionally, although the model performs well across several datasets, its ability to generalize to completely unseen sites or scanners (such as in a leave-one-site-out setting) is not addressed, which is an important consideration for real-world applicability.
- 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
The use of meta-learning with decoupled classifiers is well-justified and performs strongly across diverse datasets. Visualizations like Grad-CAM add interpretability and help build trust in the method. A brief note on computational efficiency and how the model might scale or adapt to new datasets would further strengthen the work. One important suggestion is to make the code available, as this would support reproducibility and help others build on this promising work.
- 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?
I’m recommending acceptance because this paper tackles a real and relevant problem—fMRI denoising with limited labeled data—using a creative and well-executed approach. The combination of meta-learning and decoupled classifiers is thoughtfully designed and shows strong results across diverse datasets. The experiments are solid, the writing is clear, and the visualizations add valuable insight into how the model works. While a few details could be clearer, they don’t take away from the overall strength and usefulness of the work.
- 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
We thank the reviewers for their constructive comments. Below, we address the major concerns and clarify our key design choices. We also note that the full source code, including model architectures and training configurations, will be made publicly available upon publication to ensure transparency and reproducibility.
- Rationale for Key Design Choices (#2, #3): We employ MAML [5] to train a dataset-irrelevant feature extractor that learns noise patterns across heterogeneous fMRI datasets. We simulate one-to-one domain adaptation tasks to enable robust adaptation to sparsely labeled target domains with minimal supervision. This training paradigm is particularly effective in low-label environment, where conventional approaches tend to struggle. Semi-supervised learning is essential due to the high cost of expert-labeled ICs. Leveraging few labeled samples and abundant unlabeled data, our framework reduces annotation burden while preserving performance. We incorporate wavelet-transformed ICs to enrich spectral-temporal information, complementing spatial maps and time series. These are generated using learnable wavelet kernels that automatically adapt crucial frequency bands. As shown in Fig. 2, the frequency range differs across models, although further quantitative results could not be included due to space constraints. During pretraining, we use a small number of labeled target samples to stabilize initialization and guide the model toward domain-specific denoising standards, as labeling criteria often differ across datasets. This helps avoid ambiguity in the noise/signal annotation of the target domain.
- Generalization to Unseen Sites or Scanners (#3, #4): We acknowledge the importance of generalization, but emphasize a fundamental challenge in fMRI denoising: there is no standardized labeling criterion. Labeling criteria vary across datasets, making full leave-one-site-out setups ambiguous—what constitutes “noise” is domain-specific. Therefore, a small number of labeled examples from the target domain are essential to learn the appropriate denoising standard. Nonetheless, our method is well-suited for such low-label environments due to its meta-learning and decoupled training strategy. For future work, we plan to integrate downstream tasks (e.g., brain disorder classification), enabling generalization to fully unlabeled domains through task-specific performance optimization.
- Impact of the Alignment Loss (#4): We evaluated a variant of our framework excluding L_{align} on the HCP dataset with 100% labels. The resulting F1-score was 96.45%, slightly lower than the 96.85% score with L_{align}. This demonstrates that L_{align} modestly improves domain alignment by pulling semantically similar ICs closer in the feature space.
- Comparison with One Baseline (#2): We compare with [18], the most recent and advanced deep learning-based fMRI denoising model [18] is based on prior CNN-based methods [10,13] and introduces multi-modality fusion, learnable wavelet transforms, and attention-based feature interaction. As [18] extends and outperforms earlier approaches across diverse datasets, we consider it a strong and comprehensive baseline. Comparison with [18] sufficiently demonstrates the improvements brought by our proposed framework.
- Details of the Feature Extractor (#4): Our feature extractor is based on the CNN_{sm+lwt+ts} model from [18], which processes spatial maps, wavelet-transformed images, and time series in parallel. Each stream consists of four convolutional blocks (each have two 3D/2D/1D convolution layers), using kernel size 3, stride 1, padding 1, ReLU activations, and skip connections. Attention-based D-MMTM modules [18] are applied in blocks 3 and 4, and an MLP block (block 5) follows for embedding. To support reproducibility, we will release the source code, enabling reconstruction of implementation details.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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”.
I recommend this paper for acceptance. However, based on the reviewers comments, I invite the authors to add more justification to their proposed methodology, which would thus provide new elements to judge the technical novelty.