AcNet: Aggregated channels network for automated mitosis detection

Abstract

Mitosis count is a critical predictor for invasive breast cancer grading using the Nottingham grading system. Nowadays mitotic count is mainly performed on high-power fields by pathologists manually under a microscope which is a highly tedious, time-consuming and subjective task. Therefore, it is necessary to develop automated mitosis detection methods that can save a large amount of time for pathologists and enhance the reliability of pathological examination. This paper proposes a powerful and effective novel framework named ACNet to count mitosis by aggregating auxiliary handcrafted features associated with tissue texture into CNN and jointly training neural network in an end-to-end way. Completed Local Binary Patterns (CLBP) features, Scale Invariant Feature Transform (SIFT) features and edge features are extracted and used in the classification task. In the process of network training, we expand the original training set by utilizing hard example mining, making our network focus on classification of the most difficult cases. We evaluate our ACNet by conducting experiments on the public MITOSIS dataset from MICCAI TUPAC 2016 competition and obtain state-of-the-art results.