RECOGNITION OF ULTRASOUND IMAGES USING WAVELET TRANSFORM AND ARTIFICIAL NEURAL NETWORKS Cristina Juarez-Landin, Volodymyr Ponomaryov, Jose Luis Sanchez-Ramirez ESIME-Culhuacan, National Polytechnic Institute ABSTRACT It is presented the development of two methods for recognition of ultrasound images (US) using artificial neural networks (ANN). In the first method, a neural network of the backpropagation type was used, and the second one it has been implemented a stage of extraction of characteristics applying the Wavelet transform before ANN using. The experimental results have shown the advantages and drawbacks of each a method. 1. INTRODUCTION The recognition of the images is very important in the different tasks, for example in the classification and diagnoses using medical images. Due to this reason, the implementation of algorithms of ANN is an important factor to carry out classification and recognition of the images. Theory and modeling of ANN are inspired by structure and operation of the human nervous system, where a neuron is a fundamental particle. There are four aspects that characterize a ANN: topology, learning mechanism, association type carried out among the input and output information, and the form of information representation [1]. Due to their constitution and basis, the ANN presents a great number of characteristics similar to those of a brain. This gives to them numerous advantages and therefore this technology is applied in multiple areas [2]. Recent ANN methods based on the multi-resolution analysis or multi-channels such as the filters of Gabor and Wavelet transform due to the handling of the images filtration. Other characteristic is the consideration of the filtrating of low and high frequencies due to importance in the signals and image analysis [3]. The Discrete Wavelet Transform (DWT) is applied to images providing the matrix of coefficients, known as Wavelet coefficients. Applying to an image the DWT one can obtain four types of coefficients: approaches, horizontal details, vertical and diagonal details. The approach coefficients contain most energy of the image, so is the most important, while the details have values near zero. The objective of this paper is the recognition and correct classification of ultrasound images using Wavelet Transform to obtain the principal characteristics that can be used to training the ANN. 2. METHODOLOGY 2.1. Recognition methods for ultrasound images We consider a group of eight test US images (128x128 pixels) of the patient arm that were obtained longitudinal and sequentially. We use two recognition methods, in the first one, the ANN only is applied, and in the second one, it is a realized the previous stage of extraction characteristic of the images using DWT and later the ANN is applied. The Wavelet transform of Daubechies, Symlets, Coiflets and Biorthogonal families was used in the second method with first level of decomposition, obtaining images of 64x64 pixels [4]. The figure 1 presents the decomposition scheme of an image with Wavelet transform for one level. In here, the variables M, N are the sizes of the original image. LF x is the low pass filter of the row. HF x is the high pass filter of the row. And the symbol ↓ is downsampling for elimination of the samples after apply filter. LF y is the low pass filter of the columns. HF y is the high pass filter of the columns. The final image has a half size of the original image. The component LL contains the information of low frequencies in horizontal and vertical orientations. The component HH contains the information of high frequencies in horizontal and vertical orientations. The component LH contains the information of low frequencies in horizontal orientation and high frequencies in vertical orientation. The component HL contains the information of high frequencies in horizontal orientation and low frequencies in vertical orientation. Polyphase representation of an image is an alternative approach to discrete image representation other than in the spectral and time domains. It is an efficient representation for computation. Consider the process of convolution and downsampling by 2; we compute all resulting coefficients and then cast oust half of them. 173 0-7803-9323-6/05/$20.00 ©2005 IEEE.