Improved SegMitos framework for mitosis detection in breast cancer histopathology images

Abstract

Mitotic cell counting is the strongest predictor of tumor aggressiveness in breast cancer prognosis. Since the manual annotation of mitotic cells by pathologists is extremely hard and time-consuming, automatic mitosis detection systems are highly required in pathology laboratories. In this paper, we propose a mitosis detection system inspired by the state-of-the-art SegMitos framework for which we substitute the segmentation network by the more effective DeepLabv3+ semantic segmentation model to achieve better mitosis detection performance. The improved SegMitos model consists of a downsampling path that can capture rich contextual information at multiple scales and an upsampling path that can gradually recover the image objects boundaries. Experimental results on the 2012 ICPR MITOSIS dataset and the AMIDA13 dataset demonstrate the effectiveness of our improved SegMitos system that yields better results than the original SegMitos framework and other state-of-the-art approaches with F-scores of 0.820 and 0.695 respectively.