Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis

Agostina J. Larrazabal, Nicolás Nieto, Victoria Peterson, Diego H. Milone, Enzo FerranteProceedings of the National Academy of Sciences 2020 · DOI

TL;DR

This paper demonstrates that gender imbalance in medical imaging datasets leads to biased AI classifiers for computer-aided diagnosis, with significantly worse performance for underrepresented genders. The authors urge regulatory bodies to mandate gender balance in datasets and call for the development of AI algorithms robust to such imbalances.

What Problem Does This Paper Address?

As Artificial Intelligence (AI) becomes increasingly integrated into healthcare, particularly for medical image analysis and computer-aided diagnosis (CAD), ensuring fairness and avoiding bias is critical. While efforts are made to improve diagnostic accuracy, the impact of data collection practices, specifically gender imbalance in training datasets, on AI performance has been largely overlooked. This study addresses this gap by investigating how gender imbalance in medical imaging datasets affects the performance of AI classifiers, highlighting a potential for AI systems to inherit and amplify societal biases, leading to disparities in healthcare.

Methods

The researchers employed deep neural networks, specifically a DenseNet-121 architecture (known for state-of-the-art performance in X-ray image classification), to build CAD systems. They trained these models on medical imaging datasets under various controlled gender imbalance conditions. Performance was quantified using the Area Under the Receiver Operating Characteristic Curve (AUC). To ensure robustness of their findings, they also tested two other deep learning architectures (ResNet and Inception-v3) and used two distinct X-ray datasets. The methodology involved constructing random splits of data to guarantee an equal number of pathological cases per gender within each split, thus isolating the effect of gender imbalance in the overall dataset composition.

Datasets

The study utilized two well-known publicly available X-ray image datasets:

  • NIH Chest-XRay14 dataset: This dataset contains 112,120 chest X-ray images from 30,805 patients, labeled with 14 common thoracic diseases. It has a reported gender distribution of approximately 56.5% male and 43.5% female.
  • CheXpert dataset: This dataset comprises 224,316 chest radiographs from 65,240 patients, with diagnostic information for various thoracic conditions. It has a gender distribution of roughly 60% male and 40% female.

The authors also provide access to their modified code, data splits, and additional results on GitHub, creating a reproducible benchmark for future research.

Key Insights and Ideas

  • Gender imbalance directly impacts AI performance: The study provides strong empirical evidence that imbalanced gender representation in training data leads to a significant drop in diagnostic performance for the underrepresented gender.
  • Diversity is key to generalization: A perfectly balanced dataset (50% male, 50% female) resulted in the best performance for both genders, suggesting that diversity in training data enhances the generalization capabilities of AI systems.
  • Current regulatory oversight is insufficient: The paper highlights that existing guidelines for approving AI-based medical systems, such as those from the US Food and Drug Administration, do not explicitly require gender balance or diversity in training datasets.
  • The problem extends beyond class imbalance: Unlike typical imbalanced data problems where the imbalance is between diagnostic classes, this study focuses on demographic imbalance (gender), which is often overlooked.
  • A call for robust algorithms: The authors pose an open challenge for the research community to develop novel AI algorithms that are inherently robust to gender imbalance in datasets.

Key Results

The experiments revealed a consistent decrease in performance for underrepresented genders when gender balance was not met in the training datasets. Specifically:

  • When models were trained on male-only images and tested on female images (or vice versa), a significant performance drop was observed.
  • Even with a moderate imbalance (e.g., 25% female, 75% male), the performance for the minority gender was substantially lower compared to a perfectly balanced dataset.
  • Conversely, models trained on a perfectly balanced dataset (50%/50%) achieved the best performance for both male and female test sets, outperforming models trained on highly imbalanced datasets of the same gender.

Limitations and Future Work

Limitations:

  • The study primarily focuses on gender imbalance, and while it acknowledges other potential biases (like race), it does not deeply explore their interplay with gender bias.
  • The research is conducted using X-ray images for thoracic diseases; the generalizability to other imaging modalities or medical domains is not explicitly tested.

Future Work:

  • The authors advocate for national agencies to include explicit gender balance and diversity recommendations in regulations for AI-based diagnostic systems.
  • They propose the development of novel algorithms that can achieve robustness to gender imbalance, particularly in scenarios where obtaining balanced datasets is challenging.
  • Further research is needed to explore the intersectionality of different demographic biases (e.g., gender and race) in medical AI.

Summary updated

Abstract

Artificial intelligence (AI) systems for computer-aided diagnosis and image-based screening are being adopted worldwide by medical institutions. In such a context, generating fair and unbiased classifiers becomes of paramount importance. The research community of medical image computing is making great efforts in developing more accurate algorithms to assist medical doctors in the difficult task of disease diagnosis. However, little attention is paid to the way databases are collected and how this may influence the performance of AI systems. Our study sheds light on the importance of gender balance in medical imaging datasets used to train AI systems for computer-assisted diagnosis. We provide empirical evidence supported by a large-scale study, based on three deep neural network architectures and two well-known publicly available X-ray image datasets used to diagnose various thoracic diseases under different gender imbalance conditions. We found a consistent decrease in performance for underrepresented genders when a minimum balance is not fulfilled. This raises the alarm for national agencies in charge of regulating and approving computer-assisted diagnosis systems, which should include explicit gender balance and diversity recommendations. We also establish an open problem for the academic medical image computing community which needs to be addressed by novel algorithms endowed with robustness to gender imbalance.
BibTeX
@inproceedings{larrazabal2020gender,
  author = {Agostina J. Larrazabal and Nicolás Nieto and Victoria Peterson and Diego H. Milone and Enzo Ferrante},
  title = {{Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2020},
  doi = {10.1073/pnas.1919012117},
}
Figure