3D EXTENSION OF HARALICK TEXTURE FEATURES FOR MEDICAL IMAGE ANALYSIS Ludvik Tesar Tokyo University of Agriculture and Technology, Japan email: tesar@cc.tuat.ac.jp Daniel Smutek 1st Medical Faculty Charles University Prague, Czech Republic email: smutek@cesnet.cz Akinobu Shimizu, and Hidefume Kobatake Tokyo University of Agriculture and Technology, Japan email: simiz@cc.tuat.ac.jp ABSTRACT In this paper, we propose a new approach to segmentation of 3D CT images, which is aimed at texture-based segmen- tation of organs or disease diagnosing. The extension of Haralick 2D texture feature to the 3D domain was stud- ied. Calculation of separate co-occurrence matrix for each voxel in the 3D image is proposed. The co-occurrence matrix is calculated from all voxels in a small rectangu- lar window around the voxel. This makes it possible to segment given 3D image as opposed to calculating the fea- ture for the pre-segmented regions of an image. Conse- quently, such features can be used to search for very small regions with different texture properties (like tumours). A set of abdomen CT images is used for evaluation of the proposed approach. The segmentation method we used is model-based, using Gaussian Mixture Model. EM algo- rithm is used for learning the parameters of mixture model from training data-set. KEY WORDS texture features, Haralick features, 3D image analysis, im- age segmentation, CT images, Gaussian mixture, model- based decision-making, EM algorithm 1. Introduction Texture-based image analysis using medical images is cov- ered by many theoretical as well as practical papers. We concentrate on analysis of 3D images using texture fea- tures. This area is starting to be of main interest in last years. Specificallly, we are studying new possibilities in tex- ture feature-based analysis that are opened with the avail- ability of 3D images. Texture features are often used instrument for image analysis especially for medical images [1]. Texture fea- tures are variables calculated from given image region in order to characterise the region texture. In 2D texture anal- ysis of medical images, Haralick features are often used [2, 3, 4]. In our research, we studied specific 3D extension of Haralick features in order to effectively use the three- dimensional CT scans of abdomen area. To verify usabil- ity of proposed extension of 3D Haralick texture features, we have used the features constructed, for the segmentation of abdomen area. The method used for segmentation uses model-based learning and maximum-likelihood decision- making. The model applied, is Gaussian mixture model [5]. Standard EM algorithm is used for estimation of its parameters [6]. In Section 2 we describe the specific implementation of Haralick texture features in 3D domain for 3D CT, MRI or any other 3D images available. This is the actual contri- bution of this paper. In Section 3 we describe the method how these features can be used for image segmentation. This was already presented in [5, 3, 7]. Section 4 shows the example of medical data segmentation. CT data of ab- domen were segmented using 3D features. It should be noted, that in practical application more different methods should be used and the results should be merged. Section 3.3, describes merging with different probabilistic method using Bayes formula. 2. Extension of Haralick Features to 3D Do- main Haralick 2D texture features, are statistics calculated from co-occurrence matrix. This matrix is computed from pixel intensity (gray-level or in case of CT the power inten- sity) values in a given region. The reason, why the co- occurrence matrix is so widely used in image analysis, is that it represents characteristics of the texture in a given re- gion. Haralick features are statistics defined to emphasize certain texture properties. The co-occurrence matrix con- sists of numbers, that are counts of co-occurrences of the same gray-scale colour (intensity) in two pixels separated by oriented separation vector. Number of intensity values have to be finite and relatively small, in order to have any co-occurrences in the co-occurrence matrix. Number of in- tensity value is called “quantization constant”, and denoted by q. We define original image by the 2D intensity matrix P with intensities p a,b quantized to q intensities. Region where Haralick features, are to be calculated is defined by the set R of two-component vectors. For given separation vector s = s 1 ,s 2 , the co-occurrence matrix C is defined by its components c ( s) R (i, j ) with following equation: c ( s) R (i, j )= card {u ∈ R : p u = i; p u+ s = j } card R (1) 554-026