Fig. 2: Quantitative and qualitative analysis of the H&E BE-TransMIL (ResNet50) model on the discovery test dataset. | Nature Communications

Fig. 2: Quantitative and qualitative analysis of the H&E BE-TransMIL (ResNet50) model on the discovery test dataset.

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

Fig. 2

a ROC curve (with bootstrapping for confidence intervals (CIs) between 2.5th and 97.5th percentiles) and AUROC (95% CI) at the selected operating point. b Example of a BE-positive H&E slide. Attention heatmap is heterogeneous, showing regions of high and low attentions that correspond to TFF3 staining in the adjacent TFF3 slide. Goblet cells are visible in the tiles with high attention values; tiles with low attention values do not show any goblet cells. c Example of a BE-negative H&E slide. Attention heatmap shows uniform attention without any high-attention regions; high- and low-attention tiles are without any goblet cells. Color bars along heatmaps show the range of attention values with marked mean value.

Back to article page