Image Analysis: Understanding and Mitigating Batch Effects in Histopathology

Batch effects are systematic variations, such as differences in staining or scanning protocols, that obscure true biological differences in samples. Recently trained foundation models for pathology have proven to capture morphological information effectively. However, we observe that these models may also retain irrelevant information associated with technical batch effects. While this might not harm the overall goal of improving the downstream task performance, model representations should still be robust to external distribution shifts. We advocate for systematic batch effect analysis in histopathology workflows to ensure reliable and generalizable AI models for clinical applications.