Efficient Mitosis Detection in Breast Cancer Histology Images by RCNN

Abstract

Mitotic cell detection and counting per tissue area is an important aggressiveness indicator for the invasive breast cancer. However, manual mitosis counting by pathologists is extremely labor-intensive. Several automatic mitosis detection methods have been proposed in recent years. Traditional methods using hand-crafted features suffer from large mitotic cell shape variation and the existence of many mimics with similar appearance. Pixel-wise classification working in a sliding window manner is time-consuming which hinders it from clinical application. In this work, we propose an efficient mitosis detection method in breast cancer histology images by applying modified regional convolutional neural network (RCNN). Our method achieves 0.76 in precision, 0.72 recall and 0.736 F1 score on MICCAI TUPAC 2016 datasets, outperforming all the previously published results as far as we know. F1 score of 0.585 is also achieved on ICPR 2014 mitosis dataset. TUPAC 2016 and ICPR 2014 datasets are cross validated without and with color normalization to study the generalization performance. The inference time for a 2000×2000 image is ~ 0.8 s, making our method a promising tool for clinical deployment.