Test-Time Image-to-Image Translation Ensembling Improves Out-of-Distribution Generalization in Histopathology

Histopathology whole slide images (WSIs) can reveal significant inter-hospital variability such as illumination, color or optical artifacts. These variations, caused by the use of different protocols across medical centers (staining, scanner), strongly harm algorithms generalization on unseen protocols. This motivates development new methods to limit loss generalization. In this paper, enhance robustness target protocols, we propose a test-time data augmentation based multi domain image-to-image translation. It allows project from protocol into each source before classifying them and ensembling predictions. method results in boost performances for To demonstrate its effectiveness, our has been evaluated two histopathology tasks where it outperforms conventional generalization, standard/H &E specific augmentation/normalization standard techniques. Our code is publicly available at https://gitlab.com/vitadx/articles/test-time-i2i-translation-ensembling .