computer methods and programs in biomedicine 85 ( 2 0 0 7 ) 187–195 journal homepage: www.intl.elsevierhealth.com/journals/cmpb An incremental neural network for tissue segmentation in ultrasound images Mehmet Nadir Kurnaz * ,Z¨ umray Dokur, Tamer ¨ Olmez Istanbul Technical University, Department of Electronics and Communication Engineering, 34469 Maslak, Istanbul, Turkey article info Article history: Received 29 July 2004 Received in revised form 18 October 2006 Accepted 25 October 2006 Keywords: Incremental neural network Ultrasound Image segmentation Texture analysis Feature extraction abstract This paper presents an incremental neural network (INeN) for the segmentation of tis- sues in ultrasound images. The performances of the INeN and the Kohonen network are investigated for ultrasound image segmentation. The elements of the feature vectors are individually formed by using discrete Fourier transform (DFT) and discrete cosine trans- form (DCT). The training set formed from blocks of 4 × 4 pixels (regions of interest, ROIs) on five different tissues designated by an expert is used for the training of the Kohonen net- work. The training set of the INeN is formed from randomly selected ROIs of 4 × 4 pixels in the image. Performances of both 2D-DFT and 2D-DCT are comparatively examined for the segmentation of ultrasound images. © 2006 Elsevier Ireland Ltd. All rights reserved. 1. Introduction Ultrasonography is one of the safest methods used in imag- ing human organs or their functions because of using the sound waves. However, it is difficult to determine healthy and non-healthy tissues in ultrasound images. In other words, the tissues in the ultrasound images may not be manu- ally segmented accurately and efficiently. For this reason, computer-aided segmentation of ultrasound images is quite important. The constitution of the right data space is a common prob- lem in connection with segmentation. The features that are sufficiently representative of the physical process must be searched. In the literature, it is observed that different trans- forms are used to extract desired information from biomedical images. Image intensities at one or two neighborhood of the pixel [1,2] are utilized to represent the tissues in magnetic resonance and computed tomography images. Wavelet trans- form [3–5], co-occurrence matrix [6–11], Fourier transform [12] ∗ Corresponding author. Tel.: +90 212 2853643; fax: +90 212 2853679. E-mail address: mnkurnaz@itu.edu.tr (M.N. Kurnaz). and spatial gray-level dependence matrices [13,14] are used to extract tissues in ultrasound images. Wavelet transform is used for the detection of the microcalcifications in digi- tal mammograms [15]. The second-order statistical methods include the gray-level co-occurrence matrices (GLCM) [8,16] and the gray-level run-length matrices [13]. Haralick et al. [8] proposed a set of 14 features calculated from a co-occurrence matrix, whose elements represent estimates of the probability of transitions from one gray level to another in a given direc- tion at a given inter-pixel distance. The features derived from GLCM include contrast, entropy, angular second moment, sum average, sum variance and measures of correlation. Parkkinen et al. [17] showed that GLCM can be applied on different inter- pixel distances to reveal periodicity in the texture. However, there is an inherent problem to choose the optimal inter- pixel distance in a given situation. Also, the GLCM method, in general, is not efficient since a new co-occurrence matrix needs to be calculated for every selected angle and inter-pixel distance. 0169-2607/$ – see front matter © 2006 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.cmpb.2006.10.010