Please use this identifier to cite or link to this item:
|Type:||Artigo de evento|
|Title:||Extraction And Selection Of Characteristics In Skin Lesions Images Using Wavelets In Multiresolution Analysis|
De Carvalho M.A.G.
|Abstract:||The incidence of skin cancer have increased quickly in the last few years. Although it is an illness potentially fatal, if detected and removed precociously, it has 100% of cure. The clinical diagnosis is based on the ABCD system, which is a list of parameters (asymmetry, edge, color and dimension) that had been introduced to help in the tumors diagnosis [1, 5]. This article presents a novel adaptive model based on the decomposition of a tree structure wavelet transform, for extraction and for selection of characteristics of texture in images of skin lesions. This model deals with the representation of the information contained in special frequency of skin cancer images (melanoma, benign and nomelanoma). These images can be shaped as almost-periodic signal whose dominant frequency is located in the center of the frequency band. The tree structure wavelet transform is a wavelet transform where the signal is passed though more filters than the discrete wavelet transform. In the discrete wavelet transform, each level is calculated by passing the previous approximation coefficients though a high and low pass filters . However in the tree structure wavelet decomposition, both the detail and approximation coefficients are decomposed as shown in figure 1. With the tree structure wavelet transform we are capable to work with any band of desired frequency, in order to supply the decomposition; it is not possible in conventional pyramidal wavelet transform, who favors the decomposition only of the low frequencies. Firstly, we consider the method of Chang and Kuo , which uses the calculation of the normalized average energy of each canal. When the calculated value is greater than a threshold value, the channel is decomposed. This method find the level that contains significant information and then perform the decomposition. An image is decomposed in 4 sub-images in each level of decomposition. In each sub-image f(m, n) the average energy, E, is calculated as follows E = 1/MN Σm=1M Σn=1N|f(m, n)| (1) where M and N are the pixel of the image dimensions in the x and y directions. Further decomposition of the sub-image was determined by comparing the average energy of the sub-image with the largest average energy value, Emax, in the same decomposition level. The decomposition of the sub-image is stopped if E/Emax < L, where Lisa constant threshold; otherwise, the image was decomposed further into its 4 sub-images. The main image is decomposed in four level. An example of wavelet decomposition of a sample skin lesion image with the corresponding channel designation and tree structure are given in figure 2. The decomposition of channels 1, 2 and 4, which contain most of the relevant discrimination information, is dependent on the value of the threshold constant. These tree channels are important from the classification between skin lesions according to the ABCD rule [1, 4]. It was observed that the method propose for Chang and Kuo  does not have a good performance for some classes, it is due to the facts that it has a constant threshold and uses the hypothesis that the great average energy is a good criterion of discrimination. Thus, not being possible to identify the tree structure differentiated for the class of skin lesions images. To overcome these limitations, we use the criterion of the small distribution of energy in each level with an adaptive threshold (equation 3) in each level. We compute the ratio (ρij) between the sub-image energy with the energy of the origin image described as ρij = Eij/Epj - 1 i, j = [1, 40] (2) where i indicates the sub-image and j the decomposition level. The dynamical threshold is calculated as Lj = Σi=2N Eij/N - 1 (3) where N is the total number of sub-images in each level. In our case is N = 4. If the energy ratio is greater than the dynamical threshold (ρij > Lj), the image will be decomposed. The threshold for each level is selected to show the best separation of characteristics containing excellent information for discrimination. For each class of image a tree structure is generated and for each decomposed image a energy map, called signature, is also generated (Figure 3). This signature is used to analyze the texture and to classify different images, where each decomposed image corresponds to some characteristic. Simulations were done to assure the efficiency of the method. For the simulation we used a set of 122 image, being distributed in 77 melanoma, 11 benign lesions e 34 no-melanoma lesions. Figure 4 illustrates the types of images of lesions of used skin. The method was successful in the classification of 100% for melanoma, 98% for benign lesions and 98% no-melanoma lesions. The wavelet allows to study images with different dimensions, focus in different characteristics for each image. The small wavelets coefficients can detect granularity in the images while high coefficients can represent the total trend of the image. With this research we observe a satisfactory improvement in the classification of skin cancer images.|
|Appears in Collections:||Unicamp - Artigos e Outros Documentos|
Files in This Item:
There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.