Improving Model Generalization to Out-of-Domain Data in Histopathology: Leveraging Simple Techniques for External Validation

Melanie Lubrano, Nathan Bigaud, Thalyssa Baiocco-Rodrigues, Fabien Brulport, Alexandre Filiot, David Lin

Keywords: WSI Classification, External Validation, Out-of-Domain, Calibration

Abstract: This study explores strategies to enhance the generalizability of WSI classification models across diverse centers, crucial for clinical adoption. By leveraging optimal transport and Leave-One-Center-Out training policies, we achieve a 3\% increase in the Youden Index, demonstrating their complementary roles in improving model performance. These approaches offer promising avenues for mitigating the need for extensive calibration samples, addressing challenges in external validation and ensuring robustness in clinical settings.

Submission Number: 87