Fig. 6: Results of two-step classification. The first step is the tile-level classifier. | Modern Pathology

Fig. 6: Results of two-step classification. The first step is the tile-level classifier.

From: Artificial intelligence for advance requesting of immunohistochemistry in diagnostically uncertain prostate biopsies

Fig. 6

a shows salience maps (blue for ambiguous and red for certain cases), which highlight some morphological characteristics used by the tile-level classifier to distinguish between the two classes (obtained through guided-backpropagation [30]). The second step is the slide-level IHC request decision. b, c The manifold of slide feature vectors projected onto its principal components. Each data point is the feature vector summarising tissue content for one slide. Feature vectors are labelled according to the dataset of origin (b) and the IHC request status (c).

Back to article page