Pathology AI models vulnerable to non-biological variation; new robustness measure proposed

Models for pathology foundation, which are close to being implemented clinically, face challenges from systematic non-biological variations among different centers, stemming from discrepancies in tissue processing, staining, and imaging. These inconsistencies are embedded in the models, promoting shortcut learning and diminishing their ability to generalize across various groups and institutions. The current Robustness Index (RI) measures whether local representation is influenced more by biological factors or non-biological variations; however, its count-based approach overlooks distance data. Although incorporating distance weights makes minimal impact, the fundamental issue lies in RI's pooled, fixed-neighborhood framework, which masks sample-level differences. To remedy this, the authors propose the Cross-confounder Robustness Margin (CRoMa), a sample-specific metric that assesses distances to both cross-confounder and same-confounder biological matches, offering a detailed evaluation of model robustness on an individual sample basis. This research appears on arXiv (2607.25497) as a cross announcement.