A Multi-scale Pyramid of Fully Convolutional Networks for Automatic Cell Detection

Abstract

Cell detection from the histology images is a crucial prerequisite step to narrow down the interested areas for important following analyses. While detecting cell manually is a time-consuming and labor-intensive work, automatic cell detection is in demand. However, automatic cell detection is a challenging task due to the complexity of histology images, including the variations in cell size, shape and texture. Previous methods favour handcrafted features, which lack the robustness to different cell types, while convolutional neural network based methods often suffer from the problem of maintaining spatial information and semantic information. We propose a novel network structure to leverage the balance between spatial information and semantic information on cell detection. We also design a variant of focal loss to deal with the severe imbalance between nuclei pixels and background pixels, which is a common problem in cell detection that may greatly affect the results from network. Our proposed method is evaluated on a typical colorectal histology image dataset and the result outperforms the previous state-of-the-art method (F1 score 0.834) by a large margin, achieving a new state-of-the-art (F1 score 0.889).