Automated segmentation of 3D US prostate images using statistical texture-based matching method

Abstract

A novel statistical shape model is presented for automatic and accurate segmentation of prostate boundary from 3D ultrasound (US) images, using a hierarchical texture-based matching method. This method uses three steps. First, Gabor filter banks are used to capture rotation-invariant texture features at different scales and orientations. Second, different levels of texture features are integrated by a kernel support vector machine (KSVM) to optimally differentiate the prostate from surrounding tissues. Third, a statistical shape model is hierarchically deformed to the prostate boundary by robust texture and shape matching. Experimental results test the performance of the proposed method in segmenting 3D US prostate images. © Springer-Verlag Berlin Heidelberg 2003.