An Efficient Method for Deformable Segmentation of 3D US Prostate Images

Abstract

Abstract. We previously proposed a deformable model for automatic and accurate segmentation of prostate boundary from 3D ultrasound (US) images by matching both prostate shapes and tissue textures in US images [6]. Textures were characterized by a Gabor filter bank and further classified by support vector machines (SVM), in order to discriminate the prostate boundary from the US images. However, the step of tissue texture characterization and classification is very slow, which impedes the future applications of the proposed approach in clinic applications. To overcome this limitation, we firstly implement it in a 3-level multi-resolution framework, and then replace the step of SVMbased tissue classification and boundary identification by a Zernike momentbased edge detector in both low and middle resolutions, for fast capturing boundary information. In the high resolution, the step of SVM-based tissue classification and boundary identification is still kept for more accurate segmentation. However, SVM is extremely slow for tissue classification as it usually needs a large number of support vectors to construct a complicated separation hypersurface, due to the high overlay of texture features of prostate and non-prostate tissues in US images. To increase the efficiency of SVM, a new SVM training method is designed by effectively reducing the number of support vectors. Experimental results show that the proposed method is 10 times faster than the previous one, yet without losing any segmentation accuracy.

A survey of prostate segmentation methodologies in ultrasound, magnetic resonance and computed tomography images

“…This classified feature space was subsequently used as an external force in a deformable model framework to segment the prostate. In their consequent work [81], the authors proposed to speed-up the process by using Zernike moments [82] to detect edges in low and middle resolutions and maintaining the texture classification using Gabor features and SVM. In a different way [83], the authors also proposed to reduce the number of support vectors by introducing a penalty term in the objective function of the SVM, which penalizes and rejects the outliers.…”

A random walk‐based segmentation framework for 3D ultrasound images of the prostate

“…Hybrid‐based methods can maximize the advantage of each type of information by combining multiple kinds of information, such as contour, region, and/or classification, in order to segment the prostate . Akbari et al.…”

“…trained the SVMs to measure the probability of voxels belonging to the prostate and used those classification results to define an energy function in order to drive the surface of the deformable model for the prostate segmentation . They also used a Zernike moment‐based edge detector for boundary identification and designed a new SVM training method for effectively reducing the number of support vectors . Multiatlas registration and anatomical signatures were also used for the segmentation of the prostate on ultrasound images …”

“…Hybrid-based methods can maximize the advantage of each type of information by combining multiple kinds of information, such as contour, region, and/or classification, in order to segment the prostate. [29][30][31][32][33][34][35][36][37] Akbari et al combined the statistical shape (contour information), classification information, and intensity profiles (region information) to complete the prostate segmentation. In the first step, they trained a set of wavelet support vector machines (SVM) to adaptively capture the features of the ultrasound images in order to differentiate the prostate and non-prostate tissue.…”

“…35,36 They also used a Zernike moment-based edge detector for boundary identification and designed a new SVM training method for effectively reducing the number of support vectors. 37 Multiatlas registration and anatomical signatures were also used for the segmentation of the prostate on ultrasound images. 38 Although the hybrid-based methods may be able to improve segmentation performance because of the combination of multiple information, the existing hybrid methods still suffer from two problems: (a) They depend on the population data and use the shape model or atlas obtained from the population to guide the segmentation of the prostate.…”