mlsanityData to 7 Oct 2026

The Role of Subgroup Separability in Group-Fair Medical Image Classification

arXiv 2307.02791 ↗First posted 6 Jul 2023cs.CV cs.AI cs.CYcited by 11

Not on the map. Why?

Abstract (arXiv)

We investigate performance disparities in deep classifiers. We find that the ability of classifiers to separate individuals into subgroups varies substantially across medical imaging modalities and protected characteristics; crucially, we show that this property is predictive of algorithmic bias. Through theoretical analysis and extensive empirical evaluation, we find a relationship between subgroup separability, subgroup disparities, and performance degradation when models are trained on data with systematic bias such as underdiagnosis. Our findings shed new light on the question of how models become biased, providing important insights for the development of fair medical imaging AI.

Key facts. Cited by 11 papers in the map (data through 7 Oct 2026). Builds on 17 of its 27 references that are in the map. First posted on arXiv 6 Jul 2023.

Builds on

17 of 27 references.

References · 17

  1. 2020Shortcut Learning in Deep Neural Networkscited by 1,928
  2. 2019CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparisoncited by 1,410
  3. 2017CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learningcited by 459
  4. 2018Learning Not to Learn: Training Deep Neural Networks with Biased Datacited by 255
  5. 2019Towards Fairness in Visual Recognition: Effective Strategies for Bias Mitigationcited by 252
  6. 2019Hidden Stratification Causes Clinically Meaningful Failures in Machine Learning for Medical Imagingcited by 215
  7. 2019Causality matters in medical imagingcited by 193
  8. 2021Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Datasetcited by 184
  9. 2021A Fine-Grained Analysis on Distribution Shiftcited by 171
  10. 2021Reading Race: AI Recognises Patient's Racial Identity In Medical Imagescited by 155
  11. 2018Turning a Blind Eye: Explicit Removal of Biases and Variation from Deep Neural Network Embeddingscited by 114
  12. 2022MEDFAIR: Benchmarking Fairness for Medical Imagingcited by 66
5 more
  1. 2022Leveling Down in Computer Vision: Pareto Inefficiencies in Fair Deep Classifierscited by 53
  2. 2023The Unfairness of Fair Machine Learning: Levelling down and strict egalitarianism by defaultcited by 36
  3. 2022Towards Transparency in Dermatology Image Datasets with Skin Tone Annotations by Experts, Crowds, and an Algorithmcited by 35
  4. 2020Deep Learning Applied to Chest X-Rays: Exploiting and Preventing Shortcutscited by 29
  5. 2022Potential sources of dataset bias complicate investigation of underdiagnosis by machine learning algorithmscited by 19

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2023: 3; 2024: 3; 2025: 3; 2026: 2.

11 papers.

  1. 2023FairVision: Equitable Deep Learning for Eye Disease Screening via Fair Identity Scalingcited by 13
  2. 2023No Fair Lunch: A Causal Perspective on Dataset Bias in Machine Learning for Medical Imagingcited by 11
  3. 2024Rethinking Fair Representation Learning for Performance-Sensitive Taskscited by 8
  4. 2025Understanding challenges to the interpretation of disaggregated evaluations of algorithmic fairnesscited by 6
  5. 2023Towards objective and systematic evaluation of bias in artificial intelligence for medical imagingcited by 4
  6. 2024On Biases in a UK Biobank-based Retinal Image Classification Modelcited by 4
  7. 2025Exploring the interplay of label bias with subgroup size and separability: A case study in mammographic density classificationcited by 2
  8. 2026Towards Fairness under Label Bias in Image Segmentation: Impact, Measurement and Mitigationcited by 2
  9. 2024Evaluating the Fairness of Neural Collapse in Medical Image Classificationcited by 1
  10. 2025A Primer on Causal and Statistical Dataset Biases for Fair and Robust Image Analysiscited by 1
  11. 2026Is invariance all you need for algorithmic fairness? Removing demographic information can create new biascited by 0