pplicability of mitotic figure counting by deep learning: a development and pan-cancer validation study.
Kalsnes J., Isaksen MX., Julbø F., Pradhan M., Kleppe A., De Raedt S., Skrede O-J., Torheim T., Nesheim JA., Mohn HM., Askautrud HA., Cyll K., Kildal W., Rewcastle E., Lillesand M., Kvikstad V., Janssen E., Jones R., Brustugun OT., Brennhovd B., Haug ES., Busund L-TR., Richardsen E., Andersen S., Dønnem T., Lindemann K., Kristensen G., Shepherd NA., Novelli M., Liestøl K., Kerr D., Danielsen HE., Hveem TS.
Mitotic figure counting is an established measure of cell proliferation that is included in grading systems. We developed a deep learning method for mitotic figure counting and evaluated its prognostic impact in multiple external validation datasets. The deep learning method was trained in whole slide images of tissue sections stained with haematoxylin and eosin from a publicly available breast cancer dataset where mitotic figures have been annotated by expert pathologists. The final model was externally validated according to a protocol with predefined analyses of 14 571 patient samples from 13 patient cohorts from seven different cancer types. The predefined primary analysis was univariable Cox survival analysis of the number of mitotic figures detected per mm2. Automatic mitotic figure counting correlated well with known proliferation rates, and patients with more mitotic figures per mm2 had significantly worse patient outcome in all the studied cancer types except colorectal cancer. This study demonstrates the practical potential of automated, deep learning-based mitotic figure counting, both by automating pathology work and by suggesting expanded use in more cancer types, such as prostate cancer.