Nuclei Perception Network for Pathology Image Analysis

Abstract

Nuclei segmentation is a challenge task in medical image analysis. A digital microscopic tissue image may contain hundreds or even thousands nuclear. Its morphological information provides the biological basis for the diagnosis and classification of diseases. The task requires to detect every nuclear of cells in a densely packed scene and get the segmentation of them for further pathological analysis. Nuclei segmentation can also be described as an instance segmentation task in densely packed scene. In this article, we propose a novel anchor-free dense instance segmentation framework to alleviate the issues. The network detects nuclears and segment them simultaneously. Then the nuclear segmentation mask is aggregated as nuclear instance guided by the offset map generated from the network. The network works by combining target location with pixel-by-pixel classification to distinguish crowded objects. The proposed method performs well on nuclear segmentation dataset.