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

Abstract

Diagnosis of mild cognitive impairment (MCI) and subjective cognitive decline (SCD) from fMRI functional connectivity (FC) has gained popularity, but most FC-based diagnostic models are black boxes lacking casual reasoning so they contribute little to the knowledge about FC-based neural biomarkers of cognitive decline. To enhance the explainability of diagnostic models, we propose a generative counterfactual attention-guided network (GCAN), which introduces counterfactual reasoning to recognize cognitive decline-related brain regions and then uses these regions as attention maps to boost the prediction performance of diagnostic models. Furthermore, to tackle the difficulty in the generation of highly-structured and brain-atlas constrained FC, which is essential in counterfactual reasoning, an Atlas-Aware Bidirectional Transformer (AABT) method is developed. AABT employs a bidirectional strategy to encode and decode the tokens from each network of brain atlas, thereby enhancing the generation of high-quality target label FC. In the experiments of in-house and public datasets, the generated attention maps closely resemble FC changes in the literature on neurodegenerative diseases. The diagnostic performance is also superior to baseline and SOTA models. The code is available at https://anonymous.4open.science/status/GCAN-665C.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2633_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: N/A

Link to the Code Repository

https://github.com/SXR3015/GCAN

Link to the Dataset(s)

N/A

BibTex

@InProceedings{She_GCAN_MICCAI2024,
        author = { Shen, Xiongri and Song, Zhenxi and Zhang, Zhiguo},
        title = { { GCAN: Generative Counterfactual Attention-guided Network for Explainable Cognitive Decline Diagnostics based on fMRI Functional Connectivity } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15010},
        month = {October},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The core content of this paper is to introduce a novel network structure called GCAN ( Generative Counterfactual Attention-guided Network), aimed at improving the interpretability of functional magnetic resonance imaging based functional connectivity diagnostic models for mild cognitive impairment and subjective cognitive decline.

  • Please list the main strengths of the paper; you should write about 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 GCAN model introduces counterfactual reasoning, enabling the model to identify and highlight brain regions associated with cognitive decline. This mechanism not only improves the predictive performance of the model, but also provides a deep understanding of the model’s decision-making process, thereby enhancing the interpretability of the model. (2) The AABT method adopts a bidirectional strategy to encode and decode tokens in the brain graph network, which performs well in generating highly structured and brain graph limited functional connections. This innovative design helps to more accurately reconstruct the functional connections of the target labels.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    1）Is there a correlation between the model suggested by the author and the counterfactual explanation? The latter should entail a transformation from positive to negative classes, which is currently absent in the proposed method. 2）Can the average feature map effectively capture causal information? Considering the inter-subject variability, does this averaged feature map hold meaningful implications? 3）The name “Atlas-aware Bidirectional Transformer” is very strange. Where does the term “Atlas-aware” originate from? The author just incorporated the existing knowledge of the community structure, similar to the reference 10. 4) The compared methods in this study are all based on ResNet, without any comparison to the latest methods. It is worth noting that none of these compared methods have been specifically designed for brain network analysis. 5) Figure 4 is puzzling, it is impossible to know how to obtain the counterfactual attention maps.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    The motives should be more explicit and the experiments should be conducted in a more comprehensive manner. See above for details.

  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

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

    The experiment is not enough to prove the author’s conclusion.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #2

  • Please describe the contribution of the paper

    The author proposed a generative counterfactual attention-guided network (GCAN), which introduces counterfactual reasoning to recognize cognitive decline-related brain regions and then uses these regions as attention maps to boost the prediction performance of diagnostic models.

  • Please list the main strengths of the paper; you should write about 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.

    Counterfactual reasoning is kind of interesting in this work. sufficient experiments were conducted.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    1. how does the Atlas-Aware Bidirectional Transformer perform any atlas-aware representation learning? In my understanding, only one atlas was enrolled in this work.
    2. It is challenging to identify early brain changes in AD, while the author used a pretrained classifier in GCAN. It rasise great concern whether the pretrained classifier helped in identify AD-associated regions accurately.
    3. It is weird that the diagnostic performance in HC vs. MCI is inferior to those in HC vs. SCD.
    4. Fig. 2 is confusing. What’s the output of the AABT encoder for target FC? if they are features map, how to fuse them with C_g^s?
    5. What is the L_d_c in Eq. 5? No clue of it was shown in Fig. 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    please see above.

  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

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

    The method is not rigorous enough. Small datasets and implicit figure raise great concerns in varify the fairness of this work.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #3

  • Please describe the contribution of the paper

    The main contribution of this paper is proposing a Generative Counterfactual Attention-guided Network (GCAN) for explainable cognitive decline diagnostics based on functional magnetic resonance imaging (fMRI) functional connectivity (FC) data. The GCAN introduces counterfactual reasoning to identify brain regions related to cognitive decline and uses these regions as attention maps to guide the diagnostic model.

  • Please list the main strengths of the paper; you should write about 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 GCAN architecture combines counterfactual reasoning and attention mechanisms, using counterfactual reasoning to identify brain regions related to cognitive decline and then using these regions as attention maps for the diagnostic model, enhancing explainability.
    2. The Atlas-Aware Bidirectional Transformer (AABT) is introduced for generating highly structured functional connectivity (FC) data, using a bidirectional encoding and decoding strategy that considers individual brain atlas networks to better encode and decode FC information.
    3. The proposed method is applied to the diagnosis of mild cognitive impairment (MCI) and subjective cognitive decline (SCD) using fMRI functional connectivity data, addressing a relevant and important clinical problem.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    1. Comparison to prior work on counterfactual reasoning for neuroimaging is needed. Novelty of using fMRI FC data for cognitive decline diagnosis should be highlighted compared to existing methods.
    2. Clinical feasibility and translation potential are not well discussed.
    3. Computational complexity and scalability analysis are missing.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    1. The manuscript’s claim of novelty in the GCAN architecture needs a direct comparison with the approaches of Oh et al. [7] and Ren et al. [8] to substantiate its originality in utilizing counterfactual reasoning for neuroimaging.
    2. The use of fMRI FC data in cognitive decline diagnosis is well-explored. The authors should clarify how GCAN and AABT differ from existing methods and outline their unique benefits.
    3. The practical integration of this method into clinical settings remains unclear. Discussions on its clinical applicability, including challenges and impacts on patient care, are necessary.
    4. There is a lack of detail regarding the computational demands and scalability of GCAN and AABT, which is crucial for evaluating their feasibility in handling larger or more detailed datasets.
  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    1. While the formulation of the GCAN architecture and the use of counterfactual reasoning for identifying cognitive decline-related brain regions is novel, there are similarities to prior works that are not adequately addressed or differentiated.
    2. The use of fMRI FC data for the diagnosis of cognitive decline is not entirely novel, and the authors do not clearly highlight how their approach differs from or improves upon existing methods in this domain.
    3. The clinical feasibility and potential translation of the proposed method are not well discussed, limiting the paper’s impact and relevance for clinical applications.
    4. The computational complexity and scalability of the proposed method are not addressed, which is a significant limitation given the complex nature of the GCAN and AABT architectures.
    5. The paper could benefit from a more comprehensive discussion of limitations, potential drawbacks, and future research directions.
  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #4

  • Please describe the contribution of the paper

    (1) They proposed a generative counterfactual attention-guided network (GCAN) using counterfactual reasoning to identify the cognitive decline-related brain regions. By using the identified brain regions as counterfactual attention maps, the paper innovates in enhancing the diagnostic accuracy of Mild Cognitive Impairment (MCI) and Subjective Cognitive Decline (SCD) models. (2) In the GCAN, they employed the Atlas-aware Bidirectional Transformer (AABT) to reconstruct the FC.

  • Please list the main strengths of the paper; you should write about 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 introduces an innovative counterfactual reasoning architecture aimed at identifying brain regions linked to cognitive decline. During the training stage, it constructs counterfactual attention maps specific to source labels. These maps are then applied to a masked functional connectivity (FC) matrix, which facilitates the classification of diseases associated with cognitive decline. Additionally, the implementation of a bidirectional transformer enhances the model’s capability to encode and decode the FC data, providing a more comprehensive understanding of the underlying neural interactions. (2) Well-structured experimental design, which meticulously tests various parameters of the ResNet architecture alongside different configurations of transformer heads. Additionally, The ablation study effectively demonstrates the significance of the counterfactual attention mechanism by showing a noticeable decline in performance when this component is modified or removed. (3) The validation of the GCAN model on both hospital-collected data and the publicly available (ADNI) dataset illustrates the robustness and applicability of the model across different data sources.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    (1) Applying known techniques (such as Atlas-aware configurations or counterfactual mappings) directly to a new problem area (like cognitive decline) without significant adaptation or improvement might not constitute a strong novelty. (2) The paper may not provide a comprehensive comparison with other existing models which address similar problems.

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?
    1. Documentation within the code and a detailed README that explains the structure of the codebase, the role of different scripts, and how to replicate the results for each experiment conducted in the study.
    2. Access to the datasets used in the study or clear instructions on how to obtain these datasets. If there are privacy or legal restrictions on the data, providing synthetic or anonymized datasets could be a viable alternative.
  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    1. To strengthen your claims of efficacy and innovation, a more comprehensive comparison with existing state-of-the-art models in cognitive decline diagnostics would be beneficial.
    2. Clarify the aspects of the Atlas-aware configurations and counterfactual mappings that are unique to your approach compared to existing methods. Providing explicit comparisons or discussing the incremental benefits of your modifications will help highlight your contributions more effectively.
  • 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

    Accept — should be accepted, independent of rebuttal (5)

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

    The decision to rate this paper favorably stems primarily from its innovative approach and the potential impact of its findings on the diagnosis of cognitive decline conditions like MCI and SCD. The introduction of a generative counterfactual attention-guided network and the novel use of Atlas-aware Bidirectional Transformers contribute significantly to the field. The robust experimental design and validation across multiple datasets further reinforce the validity and utility of the proposed model. However, while the paper is strong in many areas, increased comparative analysis with other models and a deeper exploration of the model’s clinical applicability would make it even stronger.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Accept — should be accepted, independent of rebuttal (5)

  • [Post rebuttal] Please justify your decision

    The authors have adequately addressed the reviewer’s comments.



Author Feedback

Thank you for the feedback. We will address the questions about the counterfactual explanation architecture, Atlas-aware, comparison methods, and clarity (Figure 2 and results) problems in the following explanations. ‘RiQj‘ means the answer to the i-th reviewer’s j-th question.

counterfactual explanation

R5Q2 and R6Q1, Q2: In existing counterfactual learning works in MRI, such as reference [7-8], they generate MRI directly, overlooking the high structural characteristics of FC, which necessitate specific design considerations. In counterfactual learning of FC, the challenge lies in generating the target label FC while maintaining strong intra-network correlations and weaker inter-network correlations. To address this challenge, we propose GCAN and AABT, which encode and reconstruct FC at atlas-network level while existing FC reconstruction works often ignore network-level generation, leading a poor performance.

R3Q1 and Q5: The model generates the target label FC, which is fundamental to the counterfactual explanation study. In this paper, the negative and positive classes can be seen as source and target labels in Figure 4. The corresponding counterfactual attention maps are generated by subtraction operation between FCs. We believe the results in Figure 4 and Table 2 can support our counterfactual study.

R3Q2: The reason why we use the average map is to supplement the disease information at the group analysis level rather than the individual level. We argue this strategy can provide more general information and can be adapted for large data analyses.

R4Q2: The pre-trained classifier provides additional information in predicting regions but will not lead to information leakage in diagnosing. Our model is trained on 1194 FC matrices when predicting AD-related regions, which is substantial given access limitations in fMRI. Inspired by pretrain strategy, we predict the regions in large datasets and transfer them to specific datasets for diagnosis.

Atlas-ware

R3Q3 and R4Q1: The reason we term it “atlas-aware” is that the hyperparameters of patch embedding and inverse embedding are determined by the atlas network. This approach effectively tackles challenges in FC counterfactual learning. When applied to a new atlas, the hyperparameters need adjustment. This stands in stark contrast to multi-atlas representation learning and community-aware methods.

Comparison

R3Q4 and R5Q1: Selection of comparison methods is guided by the prevailing SOTA approaches. Due to limitations in accessing the source code, we adopt the primary architectures of these methods and evaluate them on our dataset. Existing SOTA methods for FC classification can generally be categorized into Convolution-based (doi: 10.3389/fnins.2020.00881; 10.1002/hbm.25529) and Transformer-based methods (doi: 10.1002/hbm.26542; 10.1109/JBHI.2024.3355020). Among the Convolution family, ResNet demonstrates superior performance owing to its residual architecture. Thus, we opt for ResNet as comparison method. Transformer-based methods typically leverage direct self-attention architectures. Hence, we choose transformers with varying numbers of heads for further comparison.

Clarity (Figure 2 and results) R4Q3: In the BOLD response, MCI and SCD don’t have a hierarchical relationship. So, it’s not surprising that HC vs. MCI performance is weaker compared to HC vs. SCD. While cognitive scores may show hierarchical relationship, BOLD response amplitudes differ in regions, rather than changes in the same regions. Similar result has been reported (doi: 10.1016/j.media.2021.102248) R4Q4 and Q5: The output feature map aligns in size with the FC matrix, enabling direct addition into FC. The details of generating feature map can be seen in Figure 3. The L_d^c is the loss of image discriminator loss of C_g^s. In the latest version, it will be added nearby image discriminator in Figure 2.



Meta-Review

Meta-review #1

  • 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’

    There are four reviewers, where the two reviewers grading weakly reject are only minor issues, while as the strengths of this work as pointed by all the reviewers are significant, which can make it qualified as “accept”.

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    There are four reviewers, where the two reviewers grading weakly reject are only minor issues, while as the strengths of this work as pointed by all the reviewers are significant, which can make it qualified as “accept”.



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’

    the authors were label to clearly reply to the reviewers comments. Moreover, the overall reviewers comments are feedback is positive about the significant contribution of the research

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    the authors were label to clearly reply to the reviewers comments. Moreover, the overall reviewers comments are feedback is positive about the significant contribution of the research



back to top