Fig. 5: Generalization capabilities of the H&E BE-TransMIL (ResNet50) model on the external dataset. | Nature Communications

Fig. 5: Generalization capabilities of the H&E BE-TransMIL (ResNet50) model on the external dataset.

From: Enabling large-scale screening of Barrett’s esophagus using weakly supervised deep learning in histopathology

Fig. 5

a Stain and sample variations between the two datasets. Montages of the H&E slides in discovery (on the left) and external (on the right) datasets show lighter stain intensities and sparser tissue samples in the external dataset. b ROC curve (with bootstrapping for CIs between 2.5th and 97.5th percentiles) and AUROC (95% CI) at the selected operating point based on the discovery dataset shows competitive results despite the stain and tissue preservation variations.

Back to article page