Useful nonrobust features are ubiquitous in biomedical images

The paper's most pointed implication isn't that shortcuts exist (that's been known since Geirhos et al.) but that they are apparently ubiquitous enough across five distinct MedMNIST task types to suggest the problem is structural to how medical imaging benchmarks are constructed, not incidental to any one model architecture.

This connects directly to the cardiac diagnosis work covered the same day, 'Beyond Patient Invariance: Learning Cardiac Dynamics via Action-Conditioned JEPAs,' which identified a related failure mode: self-supervised objectives that enforce invariance end up suppressing exactly the pathological signals clinicians need. Both papers are diagnosing the same upstream problem from different angles. Standard training objectives, whether supervised or self-supervised, reward the model for whatever feature combination clears the benchmark, and in medical imaging those features are often fragile. The MedMNIST-C distribution shift results here give that concern a quantitative face.

Watch whether any MedMNIST leaderboard entrants begin reporting robustness-accuracy curves alongside top-line accuracy scores. If benchmark hosts adopt that dual reporting within the next two conference cycles, it signals the community is treating this as a methodology norm rather than a one-off finding.