Mask-Driven Mitosis Detection In Histopathology Images

Abstract

Mitotic detection and counting are the primary diagnostic factors used for cancer detection and grading. In this paper, we introduce a method of automatically obtaining masks for the cells and using the generated masks for mitotic detection. In the first stage of processing, we use the Mask R-CNN network to obtain the masks for the cells and also classify the cells with a high recall and low precision. These cells detected in the first stage are further processed in a second stage to eliminate false positives. In the second stage the cell candidates are classified as mitotic or non-mitotic using a combination of hand-crafted features and features obtained from a deep convolutional neural network pre-trained on ImageNet. We have evaluated our method on the public ICPR 2012 and ICPR 2014 mitosis datasets and obtained improved performance over fully-supervised segmentation methods and also those using a combination of hand-crafted and learned deep features. We also show that masks learned from the small ICPR 2012 dataset can be effectively transferred to the ICPR 2014 dataset, which does not provide cell mask annotations.