Classification of mitotic cells: Potentials beyond the limits of small data sets

Abstract

Tumor diagnostics are based on histopathological assessments of tissue biopsies of the suspected carcinogen region. One standard task in histopathology is counting of mitotic cells, a task that provides great potential to be improved in speed, accuracy and reproducability. The advent of deep learning methods brought a significant increase in precision of algorithmic detection methods, yet it is dependent on the availability of large amounts of data, completely capturing the natural variability in the material. Fully segmented images are provided by the MITOS dataset with 300 mitotic events. The ICPR2012 dataset provides 326 mitotic cells and in AMIDA2014 dataset, 550 mitotic cells for training and 533 for testing. In contrast to these datasets, a dataset with high number of mitotic events is missing. For this, either one of two pathologist annotated at least 10 thousand cell images for cells of the type mitosis, eosinophilic granulocyte and normal tumor cell from canine mast cell tumor whole-slide images, exceeding all publicly available data sets by approximately one order of magnitude. We tested performance using a standard CNN approach and found accuracies of up to 0.93.