Useful nonrobust features are ubiquitous in biomedical images
Coenraad Mouton, Randle Rabe, Niklas C. Koser, Nicolai Krekiehn, Christopher Hansen + 2 more
TLDR
Nonrobust features in medical images boost in-distribution accuracy but degrade out-of-distribution performance, revealing a robustness-accuracy trade-off.
Key contributions
- Nonrobust features are predictive and boost in-distribution accuracy in medical imaging.
- Models relying solely on nonrobust features achieve high accuracy on MedMNIST tasks.
- Adversarial training (robust features) improves OOD performance but reduces standard accuracy.
- Reveals a practical robustness-accuracy trade-off for medical image classification.
Why it matters
This paper highlights a critical trade-off between in-distribution accuracy and out-of-distribution robustness in medical AI. Understanding this is crucial for designing models that perform reliably in real-world clinical settings with varying data, emphasizing tailored model design.
Original Abstract
We study whether deep networks for medical imaging learn useful nonrobust features - predictive input patterns that are not human interpretable and highly susceptible to small adversarial perturbations - and how these features impact test performance. We show that models trained only on nonrobust features achieve well above chance accuracy across five MedMNIST classification tasks, confirming their predictive value in-distribution. Conversely, adversarially trained models that primarily rely on robust features sacrifice in-distribution accuracy but yield markedly better performance under controlled distribution shifts (MedMNIST-C). Overall, nonrobust features boost standard accuracy yet degrade out-of-distribution performance, revealing a practical robustness-accuracy trade-off in medical imaging classification tasks that should be tailored to the requirements of the deployment setting.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.