Fig. 4: Model attentions of BE-TransMIL models show high correspondence with TFF3 stain in BE-positive slides, and are uniform and diffused for BE-negative slides. | Nature Communications

Fig. 4: Model attentions of BE-TransMIL models show high correspondence with TFF3 stain in BE-positive slides, and are uniform and diffused for BE-negative slides.

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

Fig. 4

a Overview of TFF3 stain ratio computation. Left: slide-level attention plots for true positive slide (overlap with stain ratio, Pearson correlation r, and entropy) and true negative slide (with entropy). Right: Overlay of box and strip plots of normalized entropy of attention distributions of test dataset slides (n = 229 slides) using b H&E BE-TransMIL (ResNet50) model and (c) TFF3 BE-TransMIL (ResNet50) model. Box plots show the median (center line), 25th percentile (lower box boundary), and 75th percentile (upper box boundary), with whiskers extending to the minimum and maximum entropy values.

Back to article page