Multi-stream Faster RCNN for Mitosis Counting in Breast Cancer Images

Mitotic count is a commonly used method to assess the level of progression of breast cancer, which is now the fourth most prevalent cancer. Unfortunately, counting mitosis is a tedious and subjective task with poor reproducibility, especially for non-experts. Luckily, since the machine can read and compare more data with greater efficiency this could be the next modern technique to count mitosis. Furthermore, technological advancements in medicine have led to the increase in image data available for use in training. In this work, we propose a network constructed using a similar approach to one that has been used for image fraud detection with the segmented image map as the second stream input to Faster RCNN. This region-based detection model combines a fully convolutional Region Proposal Network to generate proposals and a classification network to classify each of these proposals as containing mitosis or not. Features from both streams are fused in the bilinear pooling layer to maintain the spatial concurrence of each. After training this model on the ICPR 2014 MITOSIS contest dataset, we received an F-measure score of 0.507, higher than both the winners score and scores from recent tests on the same data. Our method is clinically applicable, taking only around five min per ten full High Power Field slides when tested on a Quadro P6000 cloud GPU.