Automated mitosis detection with deep regression networks

Mitosis counting is one of the strongest prognostic markers for invasive breast cancer diagnosis. Clinical visual examination on histology slides by pathologists is tedious, error-prone and time-consuming. Furthermore, with the advent of whole slide imaging for high-throughput digitization, a large quantity of histology images need to be analyzed. Therefore, automated mitosis detection methods are highly demanded in clinical practice. In this paper, we proposed a deep regression network (DRN) to meet these challenges. It consisted of a downsampling path for extracting the high level information and an upsampling path for outputting the score map with original input size, thus it can be trained in an end-to-end way. In addition, we transferred knowledge learned from cross domains to mitigate the issue of insufficient medical training data. Experimental results on the benchmark dataset 2012ICPR Mitosis Detection Challenge demonstrated the efficacy of our approach, which achieved comparable or better performance than the state-of-the-art methods.