A Modified Convolutional Neural Network for Nuclei Classification in Histopathology Image

Classification of nuclei in histopathology images is an important step in pathology workflow. Although deep learning methods have been extensively employed in this task, automatic nuclei classification is still a challenging task because of the large inter-and intra-class variability as well as the serious clustered together. To address this challenge, we propose a modified automatic nuclei classification network for histopathology images. The proposed method adopts the encoding-decoding structure and leverages the segmented instance to guide the classification with the same feature maps from an encoding branch. We mainly pay attention to enhancing the feature representation. We introduce a new module composed of a modified multi-layer perceptron(MLP) module and a multi-kernel convolution(MKC) module. The MLP module focuses on mining global contextual information of nuclei and their micro-environments, and the MKC module focuses on extracting local information through different receptive fields. We validate the proposed method on two large publicly available datasets collected from different tissues. In comparison with other state-of-the-art methods, the proposed method obtains the best performance of Accuracy 88.1 and 81.6, respectively. We also verify that the proposed module can be easily incorporated into other networks for improving feature representation.