Texture enhanced Statistical Region Merging with application to automatic knee bones segmentation from CT

Abstract

Statistical Region Merging technique belongs to the portfolio of very successful image segmentation methods across diverse domains and applications. The method is based on a solid probabilistic principle and was extended in various directions to suit specific applications, including those from medical domains. In its basic implementation the technique is based on a merging criterion relying on image pixel intensities. Sufficient to segment well some natural scene images, it often deteriorates dramatically when challenging medical images are segmented. In this study we introduce a new merging criterion into the method which utilizes texture characteristic of the image. We demonstrate that the enhanced criterion allows segmentation of knee bones in CT comparable to state-of-the-art outcomes found in literature while preserving the desirable properties of the original technique.