Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Achieving Fairness Through Channel Pruning for Dermatological Disease Diagnosis

Qingpeng Kong, Ching-Hao Chiu, Dewen Zeng, Yu-Jen Chen, Tsung-Yi Ho, Jingtong hu, Yiyu Shi

TL;DR

The paper tackles fairness in dermatological disease diagnosis by introducing a Soft Nearest Neighbor Loss (SNNL)–based channel pruning framework. By computing per-channel SNNL-Fair scores, it identifies and prunes sensitive channels in the last convolutional layer to reduce discriminatory influence while preserving accuracy. The approach is validated on Fitzpatrick-17k and ISIC 2019 using VGG-11 and ResNet-18, demonstrating state-of-the-art fairness-accuracy trade-offs and compatibility with existing bias-mitigation methods, with code available at the provided repository. The method achieves notable improvements in equalized opportunity/odds and the combined FATE metric across datasets, with only a small fraction of channels pruned and a few pruning iterations needed.

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 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 state-of-the-art trade-off between accuracy and fairness. Our code is available at https://github.com/Kqp1227/Sensitive-Channel-Pruning.

Achieving Fairness Through Channel Pruning for Dermatological Disease Diagnosis

TL;DR

The paper tackles fairness in dermatological disease diagnosis by introducing a Soft Nearest Neighbor Loss (SNNL)–based channel pruning framework. By computing per-channel SNNL-Fair scores, it identifies and prunes sensitive channels in the last convolutional layer to reduce discriminatory influence while preserving accuracy. The approach is validated on Fitzpatrick-17k and ISIC 2019 using VGG-11 and ResNet-18, demonstrating state-of-the-art fairness-accuracy trade-offs and compatibility with existing bias-mitigation methods, with code available at the provided repository. The method achieves notable improvements in equalized opportunity/odds and the combined FATE metric across datasets, with only a small fraction of channels pruned and a few pruning iterations needed.

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 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 state-of-the-art trade-off between accuracy and fairness. Our code is available at https://github.com/Kqp1227/Sensitive-Channel-Pruning.
Paper Structure (18 sections, 1 equation, 3 figures, 5 tables)

This paper contains 18 sections, 1 equation, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Problem Formulation
  4. Channel Pruning for Fairness
  5. Using SNNL to reflect entanglement between different sensitive groups.
  6. Using channel pruning to achieve fairness.
  7. Pruning Recipe
  8. Experiments and Results
  9. Experiments
  10. Dataset and Models
  11. Fairness Metrics
  12. Comparison with State-of-the-art
  13. Results on Fitzpatrick-17k Dataset
  14. Results on ISIC 2019 Dataset
  15. Ablation Study
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Illustration of (1) proposed channel pruning framework and (2) distribution of initial SNNL-Fair scores, generated using the Fitzpatrick-17k dataset and VGG-11 backbone and (3) distribution of the final SNNL-Fair scores after $n$ iterations of channel pruning. $pr_c$ represents channel pruning ratio in one iteration, and $n$ is the total number of pruning and fine-tuning iteration(s) needed under the stopping criteria.
  • Figure 1: Illustration of SNNL-Fair metric calculation. Here, $t$ represents the batch index, $b$ represents the batch size, $k$ represent the depth of channel, $n_b$ represents total number of batches, and $m_b^{(t), k}$ represent the feature map at the $k$-th channel in the $t$-th batch.
  • Figure 2: FATE of $Eodd$ v.s. pruning iterations with (a) varying channel pruning ratio $pr_c$ and (b) varying layer from which the output channels are pruned, using VGG-11 on Fitzpatrick-17k dataset.