Detecting and Classifying Nuclei Using Multi-Scale Fully Convolutional Network.

Abstract

The detection and classification of nuclei play an important role in the histopathological analysis. It aims to find out the distribution of nuclei in the histopathology images for the next step of analysis and research. However, it is very challenging to detect and localize nuclei in histopathology images because the size of nuclei accounts for only a few pixels in images, making it difficult to be detected. Most automatic detection machine learning algorithms use patches, which are small pieces of images including a single cell, as training data, and then apply a sliding window strategy to detect nuclei on histopathology images. These methods require preprocessing of data set, which is a very tedious work, and it is also difficult to localize the detected results on original images. Fully convolutional network-based deep learning methods are able to take images as raw inputs, and output results of corresponding size, which makes it well suited for nuclei detection and classification task. In this study, we propose a novel multi-scale fully convolution network, named Cell Fully Convolutional Network (CFCN), with dilated convolution for fine-grained nuclei classification and localization in histology images. We trained CFCN in a typical histology image data set, and the experimental results show that CFCN outperforms the other state-of-the-art nuclei classification models, and the F1 score reaches 0.750.