Fairness: detecting bias and shortcuts

The availability of large public datasets and increased computing power have shifted the medical imaging community's focus to high-performance algorithms. However, data quality and annotations receive little attention. While ML has shown promising results in medical image diagnosis, often claiming expert-level performance, many algorithms are still hindered by issues such as bias and shortcuts. Bias primarily impacts underrepresented demographic groups based on factors such as gender or race. Shortcuts refer to spurious correlations between artifacts in images and diagnostic labels. To address this, I proposed methods for detecting shortcut learning and compiled a glossary of terms to help researchers navigate the literature [JA23b]. I conducted systematic experiments on two publicly available chest X-ray datasets that demonstrate performance degradation when images with drains are excluded.

Building on this, with colleagues at ITU, we enhance two publicly available chest X-ray datasets by adding non-expert annotations of chest tubes [CV25], using a crowdsourcing approach where data science students (non-medical experts) performed the annotations. As an output, we created the Non-Expert Annotations of Tubes in X-rays (NEATX) dataset. Further work on eye fundus images revealed models relying on irrelevant regions [ST25]. Our analysis extends beyond performance metrics, using the SHAP explainability method and embedding analysis to better understand model behavior.

Publications

[CV25] Augmenting chest x-ray datasets with non-expert annotations
Veronika Cheplygina, Cathrine Damgaard, Trine Naja Eriksen, Dovile Juodelyte, Amelia Jiménez-Sánchez
Medical Image Understanding and Analysis -- MIUA 2025 [oral]
PDF   Code   Dataset

[ST25] Mask of truth: model sensitivity to unexpected regions of medical images
Théo Sourget, Michelle Hestbek-Møller, Amelia Jiménez-Sánchez, Jack Junchi Xu, Veronika Cheplygina
Journal of Imaging Informatics in Medicine 2025
PDF   Code

[JA23b] Detecting shortcuts in medical images — a case study in chest x-rays
Amelia Jiménez-Sánchez, Dovile Juodelyte, Bethany Chamberlain, Veronika Cheplygina
International Symposium on Biomedical Imaging -- ISBI 2023
PDF   Code

Funding

  • DFF (Independent Research Council Denmark) Inge Lehmann 1134-00017B.