Shortcut learning leads to sex bias in deep learning models for photoacoustic tomography
CNN-based models trained on PAT data may engage in shortcut learning by leveraging sex-related features, leading to biased and unreliable diagnostic predictions. Addressing demographic-specific prevalence imbalances and preventing shortcut learning is critical for developing models in the medical field that are both accurate and equitable across diverse patient populations.
Authors
- Kris K. Dreher (ORCID: https://orcid.org/0000-0002-9179-9414)
- Niklas Holzwarth (ORCID: https://orcid.org/0000-0003-4440-2047)
- Werner Lang (ORCID: https://orcid.org/0000-0003-4114-7589)
- Lena Maier‐Hein (ORCID: https://orcid.org/0000-0003-4910-9368)
- Ferdinand Knieling (ORCID: https://orcid.org/0000-0002-3535-2626)
- Ulrich Rother (ORCID: https://orcid.org/0000-0002-4016-5673)
- Alexander Seitel (ORCID: https://orcid.org/0000-0002-5919-9646)
- Christoph J Bender (ORCID: https://orcid.org/0009-0009-5432-774X)
- Milenko Caranovic
- Yi Li (ORCID: https://orcid.org/0000-0003-2304-5802)
- Marcel Knopp
- Julius Kempf (ORCID: https://orcid.org/0000-0002-7036-6310)
Institutions
- German Cancer Research Center (DE)
- Friedrich-Alexander-Universität Erlangen-Nürnberg (DE)
- Heidelberg University (DE)
- University Hospital Heidelberg (DE)
- Universitätsklinikum Erlangen (DE)
- National Center for Tumor Diseases (DE)
- Heidelberg University (US)
Publication Details
- Journal
- International Journal of Computer Assisted Radiology and Surgery
- Published
- 2025-05-09
- DOI
- https://doi.org/10.1007/s11548-025-03370-9
- Citations
- 1
- Primary Topic
- Photoacoustic and Ultrasonic Imaging
- Type
- article
- Field-Weighted Citation Impact
- 0.67
Funders
- European Commission
- HORIZON EUROPE European Research Council