Weakly supervised mitosis detection using ellipse label on attention Mask R-CNN

Abstract

Automatic mitosis detection in breast histopathology cancerous tissue areas has become an important research topic recently. This paper proposed a deep learning scheme with ellipse labels for weakly supervised mitosis detection in breast histopathology images. The training labels of mitosis data are usually given only the centroid of a mitotic cell, rather than annotated every pixel of the mitosis region. The centroid labels are weak labels which are not sufficient for training a mitosis detection model. To tackle this problem, we expand the single-pixel labels to ellipse labels. We add attention mechanisms to the FPN structure of Mask R-CNN to localize and classify miotic cells. We evaluate our method on the weakly annotated 2014 ICPR MITOS-ATYPIA challenge dataset. The evaluation experiments demonstrated that our method achieved better performance compared with the baseline model and other methods, with the F-score of 0.595 in our detection task.