Applicability of mitotic figure counting by deep learning: a development and pan-cancer validation study

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Cancer diagnostics

Background:

• Mitotic figure counting is a key measure of cell proliferation used in cancer grading.
• Accurate and efficient mitotic figure counting is crucial for prognostic assessment.
• Current manual methods can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop and validate a deep learning (DL) method for automated mitotic figure counting.
• To evaluate the prognostic impact of DL-based mitotic figure counts in diverse cancer types.
• To assess the potential of DL for automating pathology workflows and expanding prognostic markers.

Main Methods:

• A DL model was trained on whole slide images of H&E-stained breast cancer tissue with expert-annotated mitotic figures.
• External validation was performed on 14,571 patient samples from 13 cohorts across seven cancer types.
• Prognostic impact was assessed using univariable Cox survival analysis based on mitotic figures per mm².

Main Results:

• The DL method demonstrated strong correlation with established proliferation rates.
• Higher mitotic figure counts per mm² were significantly associated with worse patient outcomes in most tested cancer types (excluding colorectal cancer).
• The automated counting showed practical potential for pathology automation and broader clinical application.

Conclusions:

• Deep learning-based automated mitotic figure counting is a viable and prognostic tool in oncology.
• This technology can enhance pathology efficiency and provide valuable prognostic information.
• The method shows promise for expanded use in various cancer types, including prostate cancer.