CFCN: A Multi-scale Fully Convolutional Network with Dilated Convolution for Nuclei Classification and Localization

Nuclei classification in histology images is a fundamental task in histopathological analysis. However, automated nuclei classification methods usually face problems such as unbalanced samples and significant cell morphology variances, which hinders the training of models. Moreover, many existing methods only classify individual cell patches, which are small pieces of images including a single cell. When the classification results need to be located at the corresponding position of images, the accuracy will decline rapidly, resulting in difficulties for subsequent recognition. In this paper, we propose a novel multi-scale fully convolution network, named CFCN, with dilated convolution for fine-grained nuclei classification and localization in histology images. Our network consists of encoding and decoding part. The encoding part takes cross stage partial designed network as backbone for feature extraction, and we apply cascade dilated convolution module to enlarge the receptive field. The decoding part contains transposed convolution upsampling layers, and path aggregation network is applied to fuse multi-scale feature maps. The experimental results in a typical histology image dataset show that our proposed network outperforms the other state-of-the-art nuclei classification models, and the F1 score reaches 0.750. Source code is available at https://github.com/BYSora/CFCN.