Figure 5

Image feature vectors are shown in 2-dimensional t-distributed stochastic neighbor embedded (t-SNE) plots. Top: Feature embeddings from CNN-2 are shown with (a) no prior image adaptation and (b) when image adaptation (using CNN-1) is performed prior to performing region of interest (ROI) extraction using CNN-2. Each point is an image patch within a whole slide image (WSI), colored by lab. Bottom: Feature embeddings from CNN-3, where each point represents a specimen and is colored according to ground-truth classification. (c) Specimens classified at baseline (no confidence threshold applied) are shown; (d–f) show increasing confidence thresholds (d = Level 1, e = Level 2, f = Level 3), with specimens not meeting the threshold in grey.