Improved Nuclear Segmentation on Histopathology Images Using a Combination of Deep Learning and Active Contour Model

Abstract

Automated nuclear segmentation on histopathological images is a prerequisite for a computer-aided diagnosis system. It becomes a challenging problem due to the nucleus occlusion, shape variation, and image background complexity. We present a computerized method for automatically segmenting nuclei in breast histopathology using an integration of a deep learning framework and an improved hybrid active contour (AC) model. A class of edge patches (nuclear boundary), in addition to the two usual classes - background patches and nuclei patches, are used to train a deep convolutional neural network (CNN) to provide accurate initial nuclear locations for the hybrid AC model. We devise a local-to-global scheme through incorporating the local image attributes in conjunction with region and boundary information to achieve robust nuclear segmentation. The experimental results demonstrated that the combination of CNN and AC model was able to gain improved performance in separating both isolated and overlapping nuclei.