Mitosis Detection from Pathology Images
Abstract
In the case of breast cancer, according to the Nottingham Grading System, counting mitotic cells is an important indicator of tumour diagnosis and grading. Pathologists usually manually count mitosis from histopathology images to determine the cancer grade. This is a challenging and time-consuming procedure. In most recent works, different deep neural networks have been designed to detect the suspicious cells initially and count the number of them afterwards. However, these detection approaches have certain limitations including complicated structures, the detection performance is still not satisfactory, and the need of a large number of labeled images to train a satisfied model. In this paper, we modify and improve a popular one-stage object-detection deep network to facilitate the mitotic cells detection task. Our novel improvements include using different loss functions for cells of different sizes, designing new data augmentation methods, generating prior anchor boxes with approximate sizes by using an improved clustering algorithm, and so on. We validate our deep learning model on two public benchmark datasets named Mitosis Detection in Breast Cancer Histological Images (MITOSIS). The experimental results indicate that our method achieves the competitive results on MITOSIS-2012 dataset and on the MITOSIS-2014 dataset with faster inference speed. More importantly, we design an interactive system with a "correction and relearning" pipeline so that our system can relearn from a small number of slides from a new lab and achieve satisfactory results. We design a web portal (http://ai4path.ca/#/) where this online pipeline can be easily utilized by pathologists in Western Hospital Pathology Group(WHPG) and hopefully in the future, by all pathologists in the world.