Image-encoded biological and non-biological variables may be used as shortcuts in deep learning models trained on multisite neuroimaging data.

Raissa SouzaMatthias WilmsMilton CamachoG Bruce PikeRichard CamicioliOury MonchiNils Daniel Forkert

Published in: Journal of the American Medical Informatics Association : JAMIA (2023)

The ability of DL models to exploit site-specific biases as shortcuts raises concerns about their reliability, generalization, and deployability in clinical settings.

Keyphrases

deep learning
artificial intelligence
convolutional neural network
machine learning
electronic health record
big data
resistance training