AN ENERGY-BASED SEGMENTATION OF PROSTATE FROM ULTRASOUIND IMAGES USING DOT-PATTERN SELECT CELLS Amjad Zaim 1 and Jerzy Jankun 2 1 Computer Science Department, University of Texas at Brownsville, Brownsville, Texas, USA 78520 2 Urology Research Center, Department of Urology, Medical University of Ohio, Toledo, OH 43699-0008, USA 1 amjad.zaim@utb.com, 2 jjankun@meduohio.edu ABSTRACT Automatic segmentation of prostate boundaries from Transrectal Ultrasound (TRUS) images still poses significant challenge in minimally-invasive surgical procedures. The presence of strong speckle noise and shadow artifacts limits the effectiveness of classical segmentation schemes. Several model-based and feature- approaches have been proposed for segmentation of the prostate. In this paper, we propose a new energy-based method for segmentation of ultrasound prostate images using active contour modeling guided by dot-pattern textural energy map. First, impulsive noise and speckles are reduced with median filtering and top-hat transform. Prostate features are then extracted from the filtered images using non-linear dot-pattern select operator. An elastic template shape model that incorporates a priori knowledge of the average geometric shape of the prostate boundaries as well as the energy derived from the dot- pattern feature image are utilized to search for the optimal prostate contour. A number of experiments comparing the extracted contours with manually-delineated contours validated the performance of our method. Index Terms— Prostate, Segmentation, Transrectal Ultrasound, Dot-Pattern Select Cells, Active Contour 1. INTRODUCTION Prostate cancer remains the most commonly diagnosed cancer in men and the second highest North American mortality rate among all cancers in men, surpassed only by lung cancer [1]. Modern diagnosis and treatment methods, such as needle-biopsy and brachytherapy respectively, takes into account the 2D geometric distribution of the prostate as imaged by ultrasound to map out an effective and accurate treatment plan. The relatively inexpensive and safe use of ultrasound makes it an attractive imaging modality compared to other imaging tools such as MRI and CT. To image the prostate, a cylinder shape probe, also called transrectal probe, is inserted into the rectum and rotated to scan the entire prostate capsule. The produced scans are ultimately used to reconstruct a 3-D model of the prostate [1,2]. In addition, the safety associated with ultrasound allows for real time monitoring of the prostate gland and accounts for any anatomical displacement. However, the main drawback of ultrasound stems from the presence of speckles and artifacts arising from constructive-destructive interference of the reflected waves. Ultrasound prostate images are highly corrupted with noise which prevents accurate localization of the gland. As a result, most modern treatment planning tools rely on manual outlining of the prostate; a tedious process that requires extensive labor time and comes at the expense of spatial resolution particularly when large number of 2D images are available. While many research studies have had some success in segmenting the prostate boundaries from ultrasound images with minimal human intervention, only limited progress has been reported. Researchers designed a 3D discrete active deformable model to outline the prostate using initial polygonal contours defined in a number of slices and using edge maps to drive the deformation model [3,4]. Others have developed an algorithm for detecting prostate edges as a visual guidance for the user to manually follow [5]. Statistical shape models have also been applied to segment and differentiate between the various shapes of prostates using prior knowledge of the prostate region in ultrasound images. Neural Network has also been utilized to recognize the prostate geometry from a database of prostate shapes. Gabor filtering was designed to extract prostate features and train a KSVM neural network [6]. Adaptive edge- detection methods were also employed [7]. Despite that some of these studies have reported accurate segmentation results, most still require substantial degree of user- interaction. In this paper, we propose a new template- driven approach that incorporates a priori knowledge about the average statistical shapes of the prostate to account for I  297 1424407281/07/$20.00 ©2007 IEEE ICASSP 2007