Automatic 3D Prostate Image Segmentation via Patch-based Density Constraints Clustering

Currently methods on prostate segmentation barely solve the problems about the low prostate CT contrast, high edge ambiguity, surrounding adhesion tissues and especially the tumor motion. To effectively manage those problems in prostate treatment using CT guided radiotherapy, automated segmentation needs to be performed. In this paper, an automatic 3D prostate image segmentation via Patch-based density constraints clustering (PDCC) is developed. The main contributions of this method lie in the following three strategies: 1) compared with only using pixel intensity information, Superpixel-based 3D patch includes more structure contexts to deal with low contrast problem in prostate CT images. 2) Compacting and extracting discriminative information in the each patch with 3D gray-gradient cooccurrence matrix are used to distinguish tiny texture difference between prostate and non-prostate. 3) Density constraints clustering algorithm focus on a higher density than their neighbors' points with relatively small distance to cope with two nearby organs touch together. Further, clusters are recognized regardless of their shape and of the dimensionality of the space in which they are embedded. The proposed method has been evaluated on 10 patients' prostate CT image database where each patient includes 50 treatment images, and several state-of-the-art prostate CT segmentation algorithms with various evaluation metrics have been as comparisons. Experimental results demonstrate that the proposed method achieves higher segmentation accuracy and lower average surface distance.