Abstract

Numerous studies have revealed that deep learning-based medical image classification models may exhibit bias towards specific demographic attributes, such as race, gender, and age. Existing bias mitigation methods often achieve a high level of fairness at the cost of significant accuracy degradation. In response to this challenge, we propose an innovative and adaptable Soft Nearest Neighbor Loss-based channel pruning framework, which achieves fairness through channel pruning. Traditionally, channel pruning is utilized to accelerate neural network inference. However, our work demonstrates that pruning can also be a potent tool for achieving fairness. Our key insight is that different channels in a layer contribute differently to the accuracy of different groups. By selectively pruning critical channels that lead to the accuracy difference between the privileged and unprivileged groups, we can effectively improve fairness without sacrificing accuracy significantly. Experiments conducted on two skin lesion diagnosis datasets across multiple sensitive attributes validate the effectiveness of our method in achieving a state-of-the-art trade-off between accuracy and fairness. Our code is available at https://github.com/Kqp1227/Sensitive-Channel-Pruning.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2133_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/2133_supp.pdf

Link to the Code Repository

https://github.com/Kqp1227/Sensitive-Channel-Pruning

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Kon_Achieving_MICCAI2024,
        author = { Kong, Qingpeng and Chiu, Ching-Hao and Zeng, Dewen and Chen, Yu-Jen and Ho, Tsung-Yi and Hu, Jingtong and Shi, Yiyu},
        title = { { Achieving Fairness Through Channel Pruning for Dermatological Disease Diagnosis } },
        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 paper describes the adaptation of a method for CNN channel pruning to the application of fairness in AI. The Soft Nearest Neighbour Loss is adapted to identify channels which contribute most to the performance gap between protected groups, and these are pruned followed by fine tuning of the model. This process is performed iteratively until either overall accuracy drops significantly or the fairness improvement is too small. The method is demonstrated on the task of skin lesion diagnosis using the Fitzpatrick 17k and ISIC 2019 datasets.

  • 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.
    • The paper is well written and clear
    • The method contains novelty
    • The experiments are thorough with extensive comparative evaluation and ablation tests
  • 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.
    • The novel methodological contribution with regard to the closest literature is not made clear
    • There are no statistical test in the quantitative results
  • Please rate the clarity and organization of this paper

    Excellent

  • 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.

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

    I would like to see a code release upon publication, but this is not mentioned in the paper

  • 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 paper describes an approach to mitigating biases in CNN based skin lesion diagnosis. The method adapts the previously proposed Soft Nearest Neighbour Loss (SNNL) to the problem of improving fairness. Specifically, rather than using the SNNL to measure the entanglement of channel features with class labels, it is adapted to measure the entanglement with the sensitive attribute. A proportion of channels with the lowest SNNL values are pruned from the model.

    I like the approach taken by the authors and as far as I know it is novel. The paper is very well written and clear and scientifically rigorous, with experiments on multiple datasets and multiple architectures, good comparative evaluation and ablation tests. I would like to congratulate the authors on the work they have done.

    Overall I would like to see this paper published at MICCAI, but there are a few weaknesses/questions that I would like to see addressed first. Hence, I am recommending a Weak Accept but my final recommendation will depend on the rebuttal. My concerns/questions are detailed below.

    1. My main concern is the lack of discussion in the Introduction of the most closely related literature. The Introduction is a nice general introduction to fairness in AI and dermatology in particular. But it does not review or cite previous literature that has taken a similar approach to the one described in the paper. Specifically I am thinking of work that has used pruning to achieve fairness, such as [25]. This makes it impossible for the authors to clearly state the claimed methodological novelty of the work. I request that the authors include a brief review of such works ([25] and perhaps others) and clearly state their claimed novelty in this context. I understand that there are space constraints, but to save space I wouldn’t mind some comparative results being moved from Tables 1/2 into Supplementary Materials to accommodate this change.

    2. Some of the differences in performance (accuracy or fairness) in Tables 1 and 2 are quite small. Can the authors include statistical tests for significance (e.g. against the baseline model)?

    Minor comments/questions: • Section 3.1: It is stated that the Fitzpatrick 17k dataset was used with a binary sensitive attribute (light/dark skin). Please clarify which Fitzpatrick scales were assigned to light and which to dark. • Section 3.3: The ablation study investigated the effect of pruning different (individual) layers. Couldn’t multiple layers also be pruned? • References: Please update the reference for [25] from the Arxiv paper to the MICCAI publication - https://doi.org/10.1007/978-3-031-16431-6_70

    Suggested edits: • Abstract, line 4: “achieve high-level” -> “achieve a high level” • Abstract, line 16: “achieving state-of-the-art” -> “achieving a state-of-the-art” • Section 3.1, Fairness Metrics, line 13: “set the \lambda” -> “set \lambda”

  • 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?

    I would like to accept the paper but I think a clear statement of the novel contribution of the work is a fundamental necessity for a MICCAI paper.

  • 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 #2

  • Please describe the contribution of the paper

    This paper proposes an innovative Channel Pruning method using SNNL to estimate the entanglement level between sensitive attributions and channels and remove the sensitive channels in iterations. The method has been tested in Skin lesion datasets. The results seem promising.

  • 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.

    S1: very well-structured. Nice plot, nice illustration, and nice formula. The reading experience was really good and smooth. S2: The method is novel, comprehensive experiments have been conducted, evaluation metrics makes sense. Nice ablation study on key parameters for the proposed method.

  • 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.

    W1: Lack of the details.

    • The stopping criteria. (sec 2.3, not sure if I missed it, I also didn’t find it in the experimental details)
    • number of classes for ISIC (is it multi-class classification or binary, sec 3.1)

    W2: the results do not seem to improve a lot. Compare the results from the proposed method to other baselines, it does not improve that much. Or the differences between different groups are not that obvious. It would be more clear to see the improvement if mean and std of the results is provided, rather than the result of only one run.

  • Please rate the clarity and organization of this paper

    Excellent

  • 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.

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

    I hope the code could be released on github later.

  • 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 first refer to the main strengths and weaknesses listed before. Apart from that, there are several questions I would like the authors to answer in revision or rebuttal, which are key to the implementation and usefulness of the proposed method:

    1. How robust the method is wrt. different batch splits? As mentioned in the methodology, l_sn is computed based on the samples in a batch (equation 1). How would the different data distribution in a batch influence the result? For example, batch split A: evenly split c=1 samples to all batches, batch split B: part of the batches only have samples with c=0. How would that influence your algorithm? If the batch splits matter, in practice, how would you do to avoid unwanted splits?
    2. The algorithm needs to be iterated until it reaches the stopping criteria. How long does it usually take? What are the stopping criteria? Please add the details to that. And do you need to adjust the stopping criteria according to the task, or how to choose the stopping criteria according to different tasks?
    3. In the result in Table 1, darker skin seems to have higher performance in most of the models, which is a bit anti-intuitive to me as there are more lighter color images in Fitzpatrick-17k. Could you give some explanations about this result?
  • 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?

    Novel method with solid evaluation.

  • 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 paper introduces a novel framework for addressing bias in AI-powered medical image analysis, particularly focusing on dermatological disease diagnosis. The key contribution is leveraging a modified version of SNNL measure to mitigate biases in neural networks. The paper first identifies and then selectively prunes sensitive channels that exhibit a strong association with specific demographic attributes such as skin tone or gender.

  • 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.

    The paper provides a clear and systematic methodology for channel pruning for fairness enhancement, detailing the steps involved in the process, including computation of SNNL-Fair scores and iterative pruning. The paper introduces a creative solution by repurposing SNNL to tackle bias in medical image analysis and offers a novel perspective on fairness enhancement. The authors conduct extensive experiments on diverse baselines using two dermatology datasets (Fitzpatrick-17k and ISIC 2019). The evaluation metrics are also comprehensive.

  • 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.

    While the framework demonstrates effectiveness across different datasets and models in dermatology, it is not clear if this approach can be easily generalized to cases where the sensitive attribute is non-binary.

  • Please rate the clarity and organization of this paper

    Excellent

  • 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.

  • 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

    It would be beneficial if the introduction contains a brief discussion on previous feature pruning algorithms and the way this paper is different from them.

  • 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 paper is well-written. THe methodology is novel. The authors provided extensive comparison with different baselines and datasets.

  • 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




Author Feedback

N/A




Meta-Review

Meta-review not available, early accepted paper.



back to top