A novel dataset and a two-stage deep learning method for breast cancer mitosis nuclei identification

Abstract

Mitosis nuclei counting is one of the important indicators for the pathological diagnosis and histological grade of breast cancer. With the development of deep learning methods, there have been some models for the automatic recognition of mitosis nuclei with good performance. However, due to the complex and diverse evolution stages of mitosis nuclei, automatic recognition of mitosis nuclei is very challenging, and the performance and generalization ability of the currently proposed models need to be greatly enhanced. Meanwhile, the manual annotation of images in deep learning-based model training requires experienced pathologists, which is very time-consuming and inefficient. In this paper, we propose a two-stage mitosis segmentation and classification method, named SCMitosis. Firstly, the segmentation network achieves high recall performance by incorporating the proposed depthwise separable convolution residual block and a channel-spatial attention gate, where the latter innovatively combines both channel and spatial attention mechanisms and utilizes a simple GRU unit for effective feature fusion. Then, a classification network is cascaded to improve the detection performance of mitosis nuclei further. The proposed model is verified on the pixel-level annotated ICPR 2012 dataset, which was annotated by professional pathologists, achieving the highest F1-score of 0.8687 compared with the current state-of-the-art algorithms. Additionally, the model demonstrates superior performance on the Ganzhou Municipal Hospital (GZMH) dataset, also annotated by professional pathologists, which was first released with this paper by the authors.