IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 24, NO. 1, JANUARY 2005 45 Segmentation of Kidney From Ultrasound Images Based on Texture and Shape Priors Jun Xie, Yifeng Jiang, and Hung-tat Tsui, Member, IEEE Abstract—This paper presents a novel texture and shape priors based method for kidney segmentation in ultrasound (US) images. Texture features are extracted by applying a bank of Gabor filters on test images through a two-sided convolution strategy. The tex- ture model is constructed via estimating the parameters of a set of mixtures of half-planed Gaussians using the expectation-max- imization method. Through this texture model, the texture simi- larities of areas around the segmenting curve are measured in the inside and outside regions, respectively. We also present an itera- tive segmentation framework to combine the texture measures into the parametric shape model proposed by Leventon and Faugeras. Segmentation is implemented by calculating the parameters of the shape model to minimize a novel energy function. The goal of this energy function is to partition the test image into two regions, the inside one with high texture similarity and low texture variance, and the outside one with high texture variance. The effectiveness of this method is demonstrated through experimental results on both natural images and US data compared with other image seg- mentation methods and manual segmentation. Index Terms—Image segmentation, kidney segmentation, tex- ture and shape prior, ultrasound image processing. I. INTRODUCTION I MAGE segmentation is often the first step for image analysis and is a key basis of many higher-level activities such as vi- sualization, compression, medical diagnosis and other imaging applications. The driving problem discussed in this paper is the segmentation of kidney from medical ultrasound (US) images. US imaging allows faster and more accurate procedures due to its realtime capabilities. Moreover, it is inexpensive and easy to use. The accurate detection of organs or objects from US images plays a key role in many applications. However, com- pared with other medical imaging modalities [e.g., computed tomography (CT) and magnetic resonance imaging (MRI)], US is particularly difficult to segment since the quality of the im- ages is relatively low [1]. For instance, the speckle fluctuations in the signal are proportional in magnitude to the signal strength. This property leaves US images with significant noise even in very bright regions. Moreover, because of attenuation of the probing sound wave by sound absorbing tissues, the appear- ance of most tissues change greatly such as the intensity value varies and the boundary is not always complete and prominent. Manuscript received July 19, 2004; revised September 15, 2004. The Asso- ciate Editor responsible for coordinating the review of this paper and recom- mending its publication was M. Insans. Asterisk indicates corresponding author. *J. Xie is with the Biomedical Engineering Department, Chinese Uni- versity of Hong Kong, B303, PGH3, CUHK, Shatin, Hong Kong (e-mail: jxie@ee.cuhk.edu.hk). Y. Jiang and H. Tsui are with the Biomedical Engineering Department, Chi- nese University of Hong Kong, Shatin, Hong Kong. Digital Object Identifier 10.1109/TMI.2004.837792 Bouma et al. [2] performed a study of quantitative evaluation of (semi)-automated segmentation of US images and showed that even manual segmentation of noisy US images is not straight forward. On the other hand, reliable semi-automatic segmenta- tion methods offer the potential advantage of making the mea- surement process more consistent [3]. Therefore, in this paper, we direct our research to develop a semi-automatic segmenta- tion framework, using both texture and shape priors, for kidney segmentation from noisy US images. Most image segmentation methods focus on region growing or active contours. However, the interference of speckle noise makes region growing methods [4] unreliable to classify image pixels. The active contour methods have been applied to auto- matically segment the boundaries in US images for the cortex of the brain [5], ovarian follicles [6], and for left-ventricular boundaries in echo-cardiograms [7]. Unfortunately, the basic active contour method is not adequate for our application of kidney segmentation since the tissue-tissue boundaries of kidney are relatively more difficult to localize in US images. In this paper, we present a texture and shape priors based seg- mentation framework for kidney US segmentation because we believe that a prior model of the expected anatomical structure is a significant advantage for segmenting them from US images. One of the most used methods to model shape priors is statistical modeling. Cootes et al. [8], [9] proposed the active shape models (ASM) which relies on the statistics of an object’s shape and gray-level appearances gathered from a training set of manually land-marked instances of the object. They devel- oped a parametric point distribution model for describing the segmenting curve by using linear combinations of the eigen- vectors that reflect variations from the mean shape. However, this representation does not contain any explicit information about the point connectivity. Moreover, because the techniques of automatic placement of landmarks are currently still under development (e.g., [10], [11]), landmarks are most frequently obtained manually and it is a time-consuming, error-prone and subjective work. In [12], Chen et al. represented shapes using a collection of points. They applied clustering methods instead of statistical methods to get the shape prior model which is the average shape of given curves with similar shape, but different size, orien- tation and translation. However, the similarity of shapes in this method is measured by area information which makes it highly time-consuming. Based on global-to-local registration [13] and prior region statistical properties, Rousson and Paragious [14] showed a method to recover a segmentation map in accordance with the shape prior model as well as a rigid registration between the map and the model. This method is good at accounting for local degrees of variability and local shape variations, but it con- sists of variables and, thus, is unstable. Using the Gabor 0278-0062/$20.00 © 2005 IEEE