Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Lewy Body Disease (LBD) is a common yet understudied form of dementia that imposes a significant burden on public health. It shares clinical similarities with Alzheimer’s disease (AD), as both progress through stages of normal cognition, mild cognitive impairment, and dementia. A major obstacle in LBD diagnosis is data scarcity, which limits the effectiveness of deep learning. In contrast, AD datasets are more abundant, offering potential for knowledge transfer. However, LBD and AD data are typically collected from different sites using different machines and protocols, resulting in a distinct domain shift. To effectively leverage AD data while mitigating domain shift, we propose a Transferability Aware Transformer (TAT) that adapts knowledge from AD to enhance LBD diagnosis. Our method utilizes structural connectivity (SC) derived from structural MRI as training data. Built on the attention mechanism, TAT adaptively assigns greater weights to disease-transferable features while suppressing domain-specific ones, thereby reducing domain shift and improving diagnostic accuracy with limited LBD data. The experimental results demonstrate the effectiveness of TAT. To the best of our knowledge, this is the first study to explore domain adaptation from AD to LBD under conditions of data scarcity and domain shift, providing a promising framework for domain-adaptive diagnosis of rare diseases.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5126_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YuXia_DomainAdaptive_MICCAI2025,
        author = { Yu, Xiaowei AND Zhang, Jing AND Chen, Tong AND Zhuang, Yan AND Chen, Minheng AND Cao, Chao AND Lyu, Yanjun AND Zhang, Lu AND Su, Li AND Liu, Tianming AND Zhu, Dajiang},
        title = { { Domain-Adaptive Diagnosis of Lewy Body Disease with Transferability Aware Transformer } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15966},
        month = {September},
        page = {184 -- 193}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a Transferability Aware Transformer which is a transformer-based domain adaptation method. It uses a patch-based discriminator to classify each patch as domain 1 or 2. The scores from the discriminator are used to evaluate how close patches are to each other. These scores are then used to create an adjacency matrix by multiplying its transpose to itself. The transferability matrix is used in attention calculation. Along with this, there is a global discriminator which discriminates features from the penultimate layer. The cross-entropy loss on source, global discriminator, and patch discriminator loss are added together with hyper-parameters alpha and beta.

  • 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 manuscript is clearly written, structured well, and easy to follow.

    2. Experiments were conducted thoroughly, with multiple runs per setting, and the results were appropriately reported using mean and standard deviation.

    3. The paper successfully extends the approach to address an open-set scenario, which is a valuable contribution.

  • 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 necessity of using an adjacency matrix hasn’t been adequately justified. Clarification is needed on why patch-to-patch transferability scores are essential for the proposed method.

    2. The adjacency matrix multiplication currently occurs after applying softmax, which diminishes attention scores and impacts the magnitude of outputs. Shouldn’t this multiplication be performed prior to applying softmax?

    3. It is unclear if the discriminators are trained adversarially, as the manuscript does not explicitly mention the adversarial training procedure.

    4. The method’s evaluation is limited to a single dataset/task: adapting from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset to the Lewy Body Disease (LBD) dataset. To demonstrate robustness and generalizability, additional tasks or datasets, even if similar in scope, should be included.

    5. The proposed approach has only been compared to the TVT and SSRT baselines. It would strengthen the analysis to include comparisons with standard domain adaptation methods, especially those addressing open-set problems, including non-transformer-based approaches.

    6. Performance on the novel or extra class is quite low (around 14%), indicating the model primarily focuses on source domain classes rather than effectively identifying new classes.

    7. The method exhibits high sensitivity to the probability threshold parameter, as minor adjustments (e.g., changing the threshold by 0.1) significantly degrade performance. This suggests the threshold must be carefully tuned for each dataset, reinforcing the importance of testing across multiple datasets.

    8. How were the optimal hyperparameters determined? Given only a single task was used, relying on this dataset for validation purposes could lead to incorrect or biased hyperparameter tuning.

    9. The analysis provided in the paper is limited to an ablation study. Including visualizations such as t-SNE plots would substantially enhance the understanding of the model’s behavior and feature representation.

  • 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 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

    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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Please refer to the weaknesses.

  • 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.

    The rebuttal didn’t address all of the concerns. Example: Since there is a lack of data, non-transformer architectures (CNNs) may perform better than transformers. However, the authors don’t include such results (in the main paper) or address it in the rebuttal (justify their choice).



Review #2

  • Please describe the contribution of the paper

    This paper introduces a novel method for training a transformer model to classify Lewy Body Disease (LBD) by leveraging knowledge transferred from an Alzheimer’s Disease (AD) classification task. The knowledge transfer is facilitated by a newly proposed transferability-aware transformer, designed to selectively adapt disease-specific knowledge from the AD task while discarding irrelevant domain-specific information. The proposed architecture enables effective cross-disease adaptation. Experimental results demonstrate that the method achieves robust performance in LBD classification.

  • 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 presents a novel methodology with a clear and detailed description. The proposed approach addresses an underexplored application area by transferring knowledge from a discriminator model trained to distinguish between Alzheimer’s Disease (AD) and Lewy Body Dementia (LBD). For uncertain image patches, knowledge is transferred via a self-attention mechanism. This is particularly interesting, as the method leverages the clinical similarity between the two diseases to improve performance on ambiguous regions.
    • The methodology is well-articulated and easy to follow. It is supported by relevant equations that enhance the reader’s understanding, and the implementation details are thoroughly provided.
    • The clinical relevance of this approach is notable, given that LBD is relatively understudied. Leveraging data and insights from AD to support LBD classification adds practical value to the research.
  • 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.
    • Evaluation procedures: A more comprehensive evaluation of the proposed method is needed. For instance, it remains unclear whether the model has truly learned to distinguish LBD from AD, or if it is primarily relying on features associated with AD. Clarification from the authors on this point would be valuable.
    • It would also be informative to report the overall classification accuracy across all three labels, rather than focusing solely on individual label performance.
    • The paper lacks sufficient detail on dataset preprocessing, particularly regarding how structural connectivity values were derived from the MRI data. More thorough explanations are necessary to understand and replicate the preprocessing pipeline.
  • 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
    • I recommend that the authors provide additional details regarding dataset preprocessing and related implementation aspects to enhance transparency and reproducibility
    • Further clarification of the evaluation protocol would be beneficial, particularly to validate the experimental setup and ensure the soundness of the reported results
    • For future work, it would be valuable to conduct more comprehensive experiments aimed at validating the knowledge transfer process and explicitly demonstrating how knowledge from AD classification contributes to LBD classification
  • 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?

    I think the paper and its methodology are well written, but more results are needed to show the effectiveness of the model

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    If the authors add the details mentioned in the Author Feedback, I think they would clarify my comments, so then I think the paper could be accepted. But, in my opinion, performance evaluations of LBD vs AD are actually needed.



Review #3

  • Please describe the contribution of the paper

    This paper presents an interesting framework for transferring knowledge from Alzheimer’s disease (AD) diagnosis to Lewy body dementia (LBD) diagnosis. The approach leverages foundational components of domain adaptation, particularly the use of a domain discriminator, to identify transferable features. The inclusion of an open-set decision mechanism further enhances the model’s adaptability and potential impact in clinical applications.

  • 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 addresses the important yet underexplored topic of Lewy Body Dementia (LBD) diagnosis. Given the scarcity of data in this domain, the proposed method has the potential to make a meaningful contribution by enabling knowledge transfer from related tasks.
    2. The paper is clearly written and well-organized, making it easy to follow and understand.
  • 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 the open-set adaptation component, the paper introduces a threshold-based approach to handle the third category for LBD diagnosis. However, given that some training data from the LBD domain is assumed to be available, it would be helpful to clarify why a simple shallow classifier was not considered as an alternative. Such an approach might offer improved performance and potentially simplify the decision process.

  • 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 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 proposed method has strong potential to make a meaningful contribution by facilitating knowledge transfer from AD diagnosis to LBD diagnosis, an area that remains underexplored in current research.

  • 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 rebuttal addresses my questions well. I am inclined to recommend accept.



Author Feedback

We appreciate the valuable comments and the recognition of the strengths in our paper: (1) The importance and clinical relevance of the topic on Lewy Body Disease research (R1, R3) (2) Clear writing and well-organized presentation, easy to follow (R1, R2, R3) (3) Thorough experimental evaluation/Results demonstrate robustness (R2, R3)

Code and data will be released upon acceptance.

We summarize and address the concerns:

Reviewer 1:

In our setting, labels for the target domain (LBD) are not accessible during training. In the training, the source domain (AD) is supervised, while the target domain (LBD) is unsupervised. Due to label discrepancy and the lack of supervision in the target domain, we cannot directly train a multi-class classifier to distinguish among CN, MCI, and LBD. Instead, we adopt a threshold-based entropy approach to identify samples that do not confidently match source classes, addressing this open-set scenario.

Reviewer 2:

  1. Why patch-to-patch score? Our method builds on the Vision Transformer (ViT), which requires input to be divided into patches. However, ViT does not model explicit relationships between patches beyond positional embeddings. We introduce a learnable adjacency matrix that encodes patch-to-patch transferability scores, enabling the model to focus on semantically meaningful interactions across patches.

  2. Clarification on adjacency matrix ordering. As shown in Eq. (7), the adjacency matrix multiplication is indeed performed before applying the softmax, as the reviewer suggested. Eq. (4), which the reviewer may be referring to, only describes the final transformer layer.

  3. Adversarial training procedure. The discriminators are trained adversarially alongside the rest of the model using an adversarial loss. The discriminator design follows the established approach in [1], which has proven effective in similar settings. [1] TVT: Transferable Vision Transformer for Unsupervised Domain Adaptation, WACV 2023.

  4. Additional datasets, comparison models/architecture, ablation study, and tune optimal parameters. Our method is tailored for LBD, a rare disease with extremely limited labeled data. We provide ablation studies in Tables 2 and 3, and compare against strong baselines including TVT [1] and SSRT [2]. Extension to other datasets and tasks will be explored in future work. Note: MICCAI policy prohibits inclusion of additional experiment results. [2] Safe Self-Refinement for Transformer-based Domain Adaptation, CVPR 2022.

  5. Performance on LBD. Unlike datasets with abundant labels, LBD data is scarce. Existing methods like TVT [1] and SSRT [2] perform poorly under these conditions. In contrast, our model achieves higher accuracy on the unseen LBD class, establishing a promising baseline for LBD disease research.

  6. Threshold sensitivity. The sensitivity to the threshold arises from LBD data scarcity and class imbalance, a common phenomenon in rare disease analysis. As more LBD data becomes available, this issue will be effectively mitigated.

Reviewer 3:

  1. Transferable knowledge and model validation. The model’s ability to identify the LBD class without label supervision indicates it has learned transferable features beyond AD-specific traits. We agree that further analysis is valuable and will include feature attribution and clustering visualizations in the final version.

  2. Overall accuracy. With optimal hyperparameters, the model achieves 72.3 ± 5.8% overall accuracy. Importantly, the model is trained without LBD labels, reflecting real-world limitations and the importance of domain adaptation from AD.

  3. Data preprocessing / Structural connectivity derivation. We follow the procedure in [3] to derive structural connectivity from MRI. We will provide more detailed descriptions in the revision and release our preprocessing code and data upon acceptance. [3] Predicting brain structural network using functional connectivity, Medical Image Analysis 2022.



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.

    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 #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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top