Preventing Shortcut Learning in Deep Neural Networks for Medical Applications

Project Level: Honours, PhD

Shortcuts are decision rules that perform well on standard benchmarks but fail to transfer to more challenging testing conditions, such as real-world scenarios. This can happen when the algorithm learns to use spurious correlations to perform the prediction task. For example, an algorithm learned how to detect pneumonia from chest X-rays by detecting which hospital the scan came from using the labels on the photos [1]. In this project, we develop strategies to prevent shortcut learning in medical applications [2]. 

[1] Zech, J. R. et al. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLoS Med. 15, e1002683 (2018).
https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1002683

[2] Geirhos, R., Jacobsen, JH., Michaelis, C. et al. Shortcut learning in deep neural networks. Nat Mach Intell 2, 665–673 (2020).
https://doi.org/10.1038/s42256-020-00257-z