Fig. 6

Inference results before (top images a, b, and c) and after (bottom images d, e, and f) training with background ROIs. Examples are shown of the models learning new features and what to ignore as background objects that are not Pre-NFTs/iNFTs. a and d Models learn to ignore corpora amylacea. In the training dataset there were no examples of these objects and thus were originally predicted as NFTs. b and e Edges or vessels are also often predicted as Pre-NFTs or iNFTs since the model was never exposed to these during training, but can learn to ignore these as background. c and f Folded tissue was also a common mistake as it provided a sudden darker shade compared to background, oftentimes having edges that look of NFT shape. However, new models learn to ignore these after seeing examples of folded tissue. Red boxes: iNFTs, blue boxes: Pre-NFTs. Inference Mistake and Corrections for WSIs