An artificial intelligence system that can differentiate cancer from healthy tissue is reported this week in Nature Medicine. This algorithm could help pathologists exclude up to 75% of uninformative tissue samples while retaining 100% sensitivity, thus aiding in diagnosis and accelerating routine clinical practice in cancer centres.
The use of digital pathology systems designed to improve care of patients with cancer has been hindered by a lack of models with clinical diagnostic accuracy and the ability to cope with the large number of cases that are routinely processed in cancer hospitals.
Thomas Fuchs and colleagues compiled a large real-world dataset of more than 44,000 tissue slides from over 15,000 patients diagnosed with prostate, skin and breast cancers. They built a deep learning model able to recognize tumour cells in histological preparations without need for manual annotation by a pathologist. They found that the model could diagnose these tumour types with clinical-grade performance, even when irregularities in the samples were present, such as air bubbles, knife slices or folds.
The authors suggest that this approach could streamline the workflow of pathologists, allowing them to focus their efforts on reviewing informative tissue sections that contain tumour tissue.