Wavelet Transform in Biomedical Image Segmentation and Classification Aleˇ s Proch´ azka and Oldˇ rich Vyˇ sata Institute of Chemical Technology in Prague Department of Computing and Control Engineering Technick´ a 5, 160 00 Prague 6, Czech Republic Emails: A.Prochazka@ieee.org, vysata@neurol.cz Magdal´ ena Kaˇ sparov´ a and Ta ˇ tjana Dost´ alov´ a 2nd Medical Faculty, Charles University Department of Paediatric Stomatology V ´ Uvalu 84, 150 06 Prague 5, Czech Republic {Magdalena.Kasparova, Tatjana.Dostalova}@fnmotol.cz Abstract—The contribution is devoted to the study of image segmentation and texture analysis to find image features invariant to image components rotation and translation. The main part of the paper presents the principle of Radon transform and its use in combination with the wavelet transform to find features minimizing their variance due to image components rotation. Proposed methods have been verified for simulated structures and then used for analysis of biomedical images including magnetic resonance images of the brain and orthodontic images. The goal of image processing included in all cases (i) segmentation of selected biomedical objects and (ii) detection of their features. I. I NTRODUCTION Image components classification and object recognition form a very important research area [1], [2], [3] closely related to feature extraction invariant to scaling, rotation, translation and illumination. There are various methods contributing to solution of these problems [4], [5], [6] published recently. (a) MR IMAGE (b) RIDGE LINES (c) SEGMENT Fig. 1. An example of MR image segmentation presenting (a) original MR area, (b) ridge lines resulting from its watershed segmentation, and (c) a selected segment The paper is devoted to the study of the effect of image de- noising and enhancement [7], [8], [9], [10] in connection with biomedical magnetic resonance (MR) image segmentation and classification. The main part of the paper presents a discussion of the use of Radon and wavelet transforms to extract image components features invariant to their rotation and translation [11], [12], [13]. The proposed method has been verified for simulated images consisting of segments having the same texture but a different angle of their rotation. Using the rotation-invariant texture analysis technique these features, corresponding to individual segments corrupted by noise, have been evaluated and their standard deviation has been used as a measure of the efficiency of the whole method. (a) GIVEN IMAGE (b) GRADIENT MAGNITUDE IMAGE SEGMENTATION - 1 Fig. 2. An example of an orthodontic image segmentation presenting (a) orig- inal orthodontic image, (b) gradient image enhancement, and (c) segmentation resuls Resulting algorithms have been verified both for simulated and real biomedical images. Similar approach has been applied for magnetic resonance images [14] and orthodontic images [15], [16] to find their features for the subsequent use of self- organizing neural networks [17] to classify image components and to enable more detail diagnosis. II. WATERSHED TRANSFORM IN I MAGE SEGMENTATION While the feature based segmentation can be very efficient [18] it is possible to detect image components using the distance and watershed transforms [17] followed by image ridge lines estimation in many cases. The proposed algorithm combines feature estimation based both upon the analysis of texture structure and image segment boundary signal. The watershed segmentation [19], [20] used for the subse- quent image analysis with a selected result presented in Fig. 1, Fig. 2 and Fig. 3 include • image thresholding to find its black and white form • distance and watershed transforms use to find ridge lines • extraction of the segment boundary signal The proposed algorithm combines these methods with the gradient image processing presented in Fig. 2 to enhance image objects boundaries. Image segment features are then estimated and stored in the pattern matrix for their classification. The proposed method uses feature extraction [10] by analysis of both boundary signal and image texture combining the Radon and wavelet transforms to eliminate their dependence upon image rotation and translation. The problems of image oversegmentation has been reduced by the initial image de-noising.