This paper introduces PathoROB, the first systematic robustness benchmark for evaluating Foundation Models (FMs) used in digital pathology. The study investigates how susceptible these large-scale AI models are to learning non-biological technical features (like scanner hardware or staining variations) rather than focusing exclusively on biological signals. The authors present three novel metrics, including the robustness index, to quantify these deficits, demonstrating that limited robustness can lead to major diagnostic errors ("Clever Hans" learning) in downstream clinical tasks. Furthermore, the paper explores post-hoc robustification techniques like stain normalization and ComBat batch correction to mitigate these issues without requiring costly FM retraining, ultimately establishing that robustness evaluation is essential for safe clinical adoption and should be a core design principle for future pathology FM development.

References:

• Kömen J, de Jong E D, Hense J, et al. Towards Robust Foundation Models for Digital Pathology[J]. arXiv preprint arXiv:2507.17845, 2025.