Parallel YOLO-based Model for Real-time Mitosis Counting

Abstract

It is estimated that breast cancer incidences will increase by more than 50% by 2030 from 2011. Mitosis counting is one of the most commonly used methods of assessing the level of progression, and is a routine task for every patient diagnosed with invasive cancer. Although mitotic count is the strongest prognostic value, it is a tedious and subjective task with poor reproducibility, especially for non-experts. Object detection networks such as Faster RCNN have recently been adapted to medical applications to automatically localize regions of interest better than a CNN alone. However, the speed and accuracy of newer state-of-the-art models such as YOLO are now leaders in object detection, which had yet be applied to mitosis counting. Moreover, combining results of multiple YOLO versions run in parallel and increasing the size of the data in a way that is appropriate for the specific task are some of the other methods can be used to further improve the score overall. Using these techniques the highest F-scores of 0.95 and 0.96 on the MITOS-ATYPIA 2014 challenge and MITOS-ATYPIA 2012 challenge mitosis counting datasets are achieved, respectively.