Abstract
The automated histopathology using whole slide images (WSIs) and deep learning-based algorithms (AI pathology) should be useful for efficient toxicity detection of candidate compounds. However, there are few reports on the prediction accuracy when dealing with a large number of findings or variations in hematoxylin and eosin staining of specimens. Additionally, while AI pathology generally makes predictions on small-sized images obtained by cropping WSIs, toxicological pathologists make comprehensive diagnoses at the whole slide level. Here, investigation was conducted to determine the extent to which the automated histopathology using AI pathology is possible. First, an AI pathology models using EfficientnetV2 were constructed for 18 findings using images obtained from WSIs, which were stored in Open TG-GATEs and were scanned from rat liver specimens created in multiple facilities with staining variations. The models achieved an average F1 score exceeding 0.946. Furthermore, when this model was used to automatically diagnose 11,299 rat liver WSIs stored in Open TG-GATES, it showed a high level of consistency with the diagnoses of pathologists (sensitivity, 0.913; recall, 0.870). Based on these results that our AI pathology models could predict generally consistent diagnoses with pathologists in numerous WSIs showing diverse staining, it was suggested that an AI pathology model capable of automating histopathological examination of rat livers can be constructed.
