UNSUPERVISED SHAPE PRIOR MODELING FOR CELL SEGMENTATION IN NEUROENDOCRINE TUMOR.

Automated and accurate cell segmentation provides support for many quantitative analyses on digitized neuroendocrine tumor (NET) images. It is a challenging task due to complex variations of cell characteristics. In this paper, we incorporate unsupervised shape priors into an efficient repulsive deformable model for automated cell segmentation on NET images. Unlike other supervised learning based shape models, which usually require a large number of annotated data for training, the proposed algorithm is an unsupervised approach that applies group similarity to shape constraints to avoid any labor intensive annotation. The algorithm is extensively tested on 51 NET images, and the comparative experiments with the state of the arts demonstrate the superior performance of this method using an unsupervised shape model.