PAN AMERICAN HEALTH CARE EXCHANGES PAHCE. CONFERENCE, WORKSHOPS, AND EXHIBITS. COOPERATION / LINKAGES.

. INTERCAMBIOS DE CUIDADO MDICO PANAMERICANOS. CONFERENCIA, TALLERES Y EXHIBICIONES. COOPERACIN / ENLACES.

Automatic Segmentation of Digital Images Applied in Cardiac Medical Images

F. A. Peres1, F. R. Oliveira1, L. A. Neves1,2,3, M. F. Godoy3
2

Faculdade de Tecnologia de So Jos do Rio Preto, So Jos do Rio Preto, SP, Brasil Universidade Estadual Paulista, Departamento de Estatstica, Matemtica Aplicada e Computao, Rio Claro, SP, Brasil 3 Faculdade de Medicina de So Jos do Rio Preto, So Jos do Rio Preto, SP, Brasil Email: fabin.peres@hotmail.com, franciscoliveira7@hotmail.com threshold takes a subjective criterion of a human operator. Several methods had been proposed to do this automatically based on different criteria in the image, such as those proposed [3], [4], [5], [6], [7] e [8]. However, in many cases is not achieved a threshold that provides a good segmentation of the entire image. For such situations are applied techniques of variables and multilevel thresholding based on analytical studies, using the parameters of statistical distribution of gray levels, or graphics, using the histogram display the gray level image. For this, [9] proposed a methodology where the algorithm automatically gets the multilevel threshold, by the histogram analysis. The method finds the histogram valleys, which are the places where are concentrated the thresholds and therefore the subdivision of the image. However the method proves effective in cases where the image and the histogram are well defined, for cases where the image is not presented optimally, with noise, distortion and nonstandardized histograms, the method does not produce an effective threshold that identifies the objects in the image quality. In this context, the paper presents an improved version of the model of multilevel automatic thresholding described in [9], considering the group histogram quantization, analysis of the histogram slope percentage and calculation of maximum entropy to define the threshold. These improvements prevent the identification of not significant thresholds and allows more control of the technique during the step of feature extraction in artificial vision systems.

Abstract The digital image processing has been applied in several areas, especially where it is necessary use tools for feature extraction and to get patterns of the studied images. In an initial stage, the segmentation is used to separate the image in parts that represents a interest object, that may be used in a specific study. There are several methods that intends to perform such task, but is difficult to find a method that can easily adapt to different type of images, that often are very complex or specific. To resolve this problem, this project aims to presents a adaptable segmentation method, that can be applied to different type of images, providing an better segmentation. The proposed method is based in a model of automatic multilevel thresholding and considers techniques of group histogram quantization, analysis of the histogram slope percentage and calculation of maximum entropy to define the threshold. The technique was applied to segment the cell core and potential rejection of tissue in myocardial images of biopsies from cardiac transplant. The results are significant in comparison with those provided by one of the best known segmentation methods available in the literature. Keywords cardiac Imagens; segmentation; thresholding

I. INTRODUCTION The biological vision system is one of the most important means of exploration of the world to the humans, performing complex tasks with great ease such as analysis, interpretation, recognition and pattern classification. For this reason many studies attempt to produce artificial vision systems with the same efficiency of the biological system. This task is still highly complex, mainly to implement one of the most obvious problem, the quantification and qualification of informations represented in many different fields, such as intensity of gray level, edges, contours and texture. These attributes are naturally sought by the human visual system when the measured signal is an image [1]. One possibility to represent an artificial vision system efficient is to use appropriate methods of segmentation, considered as a first step for analyzing an image, it allows to separate the objects in parts, according to some criterion of uniformity [2]. For high quality segmentation systems, digital image processing is used in a primary stage of thresholding to separate the object of the rest of the image. The thresholding consist in to identify in an image, a threshold of intensity in which the object distinguish better of the back of the image, and in most cases, the choice of
MARCH 15 - 19, 2010, LIMA, PERU 978-1-4244-6294-0/10/$26.00 2010 IEEE 38

II. METHODOLOGY The technique was based on a model of multilevel thresholding proposed in [9]. One of the first improvements made considers the segmentation of the histogram in specific groups. The group size is a user parameter, and if the given value is 1, the analytical process analyzes each gray level of the histogram. One example is to consider a group of 10 values and 256 gray levels. The first group will include the gray levels from 0 to 9, the second from 10 to 19 and so on. In Figure 1 we have a graphic demonstration of as will be the scenario described, and the vertical lines indicating the separate group. The identification of the histogram valleys is of paramount importance, because in these valleys, the
ISBN 978-1-4244-6294-0 IEEE CATALOG NUMBER CFP1018G-ART

PAN AMERICAN HEALTH CARE EXCHANGES PAHCE. CONFERENCE, WORKSHOPS, AND EXHIBITS. COOPERATION / LINKAGES. INTERCAMBIOS DE CUIDADO MDICO PANAMERICANOS. CONFERENCIA, TALLERES Y EXHIBICIONES. COOPERACIN / ENLACES.

thresholds are concentrated, and therefore the division of classes. Therefore, another improvement realized is the ability of the algorithm to automatically identify these valleys, through the transition of histogram signal values. This is accomplished by comparing the first value of the group, which was determined in the segmentation of the histogram, with the last of the same group. If the first value is less, means that the histogram values are increasing, and the sign becomes positive. Otherwise, the histogram values are declining, and the sign, therefore, be negative. The next step is to check the sign of the next group. Whenever there is a transition from a negative to positive, there is indication of a valley. Figure 2 illustrates this process. Once you found the first valley, passes to the next step, the analysis of the slope percentage. In step for analysis of slope percentage it is possible that occur divergences in cases where there is a uniform histogram, without valleys or valleys insignificant, with little variance, as seen in the first valley of Figure 2. For the type of image analyzed, we identified that an effective threshold would be near to the base of a group that has a considerable slope percentage. This minimum percentage of slope is set by the user, who may adjust according to the type of image. This approach involves calculating the slope achieved by the difference of the average of the last three values of the group with the first three. If this difference is greater than the parameter set, the scanning of the slope percentage is interrupted and go to the step of identifying the threshold. If the difference is smaller, the histogram scan continues until the difference of means is greater. This strategy avoids identifying groups and trace thresholds for classes with little variation, separating objects with common characteristics.

Fig. 2 . Exemplification of signal transition in the gray levels of the histogram.

Having established the group with possible valley, the threshold identification can happen in two different ways, by seeking the lowest point or the calculation of maximum entropy. The strategy at the lowest point compares analytically the gray levels values of the group to identify what has fewer occurrences. This is stored and identified as a threshold. The thresholding based on maximum entropy is achieved from probabilistic calculations. In this context we consider an image as a result of a random process, where the probability pi corresponds to the probability of a pixel of the image taking a value of intensity i (i = 1, .., n) [10], as shown in (1) and (2).

 pi * log pi
i 1

(1) (2)

pi

ni N

where H is the entropy of the image, n is the total number of outputs, i.e., the number of gray levels of image; pi is the probability of gray level i to be found in the image; ni represents the number of pixels with intensity i and N is the total number of pixels in the image. The gray level of the group indicated with the highest entropy is identified as a threshold. Completed this phase, the process continues with the analysis of the other groups. To test the technique, we used myocardial images of biopsies from heart transplant. The choice of this type of image to processing is justified by the groups of existing objects, cell core, fibrous tissue, muscle and tissue rejection. 120 images were processed and the results were compared with those provided by the technique presented in [3].

Fig. 1 . Exemplification of the histogram segmented in groups of analysis.

MARCH 15 - 19, 2010, LIMA, PERU

39

ISBN 978-1-4244-6294-0 IEEE CATALOG NUMBER CFP1018G-ART

PAN AMERICAN HEALTH CARE EXCHANGES PAHCE. CONFERENCE, WORKSHOPS, AND EXHIBITS. COOPERATION / LINKAGES. INTERCAMBIOS DE CUIDADO MDICO PANAMERICANOS. CONFERENCIA, TALLERES Y EXHIBICIONES. COOPERACIN / ENLACES.

III. RESULTS In the studied images we used four different input parameters, the image format to be processed, the size of the histogram division group, the filter size, and the percentage of slope to be used for identify thresholds. To demonstrate the method and the results, was used the Figure 3, having as input parameters: the image format = jpg, size of the histogram division group = 10, filter size = 5 and slope percentage = 35%. These parameters can be adjusted to the type of image that pretends analyze. To analysis the results, are compared three items, namely: the thresholds outlined, that are the points that are the division of the objects or classes, the average gray level for each class, and the total points of each class. These data show us how significant is the object to the whole image. The results and the differences between the techniques, comparing the thresholds, the average gray levels and the total points are shown in Tables I, II and III, respectively.

TABLE III COMPARISON OF TOTAL POINTS OBTAINED BY THE LOWEST POINT TECHNIQUE AND BY MAXIMUM ENTROPY Class 1 11.721 13.823 0,44% Total Points Class 2 Class 3 408.926 59.353 408.421 57.756 -0,11% -0,33%

Lowest Point Maximum Entropy Difference

As shown in Table I, the thresholds difference is small, 3,51% for the first threshold. The differences between the methods are also insignificant in comparison of the average gray level and total points of each class. In the average gray levels, the differece ranged from -0,11% to 2,59%. In total points, the difference ranged from -0.33% to 0.44%. These differences not present a significant visual result to human vision. For this reason, we will show the final results of the segmentation process with the proposed method, using the technique of maximum entropy, Figures 4 to 6.

Fig. 3 . Example of myocardial images obtained with biopsies of a transplanted heart patient.

Fig. 4 . Filtered histogram of the image shown in Fig 3.

TABLE I COMPARISON OF THRESHOLDS (GRAY LEVELS) OBTAINED BY THE LOWEST POINT TECHNIQUE AND BY MAXIMUM ENTROPY Lowest Point Maximum Entropy Difference Thresholds (Average) 61 236 70 240 3,51% 1,56%

TABLE II COMPARISON OF AVERAGE GRAY LEVELS OBTAINED BY THE LOWEST POINT TECHNIQUE AND BY MAXIMUM ENTROPY Average Gray Levels Class 1 Class 2 Class 3 34,49 165,20 243,82 41,12 165,99 243,55 2,59% 0,31% -0,11%
40

Lowest Point Maximum Entropy Difference

Fig. 5 . Result of image processing shown in Fig 3, showing the cell core or tissue rejection obtained with the maximum entropy.
ISBN 978-1-4244-6294-0 IEEE CATALOG NUMBER CFP1018G-ART

MARCH 15 - 19, 2010, LIMA, PERU

PAN AMERICAN HEALTH CARE EXCHANGES PAHCE. CONFERENCE, WORKSHOPS, AND EXHIBITS. COOPERATION / LINKAGES. INTERCAMBIOS DE CUIDADO MDICO PANAMERICANOS. CONFERENCIA, TALLERES Y EXHIBICIONES. COOPERACIN / ENLACES.

TABLE VI COMPARISON OF TOTAL POINTS OBTAINED BY MAXIMUM ENTROPY AND BY OTSUS METHOD (1978), [3]. Total Points Class 1 Class 2 Class 3 52.160 277.376 150.464 13.823 408.421 57.756 7,99% -27,30% 19,31%

Otsu (1978) Maximum Entropy Difference

Fig. 6 . Result of image processing shown in Fig 3, showing cardiac muscle.

For comparison of the segmented regions, beyond the strategy of maximum entropy, we compared the results with those provided by the Otsu's method, [3]. For this, we used the same criteria and parameters described in the comparison between the lowest point and maximum entropy, which are the thresholds, the average gray level and the total points for each. In Tables IV, V and VI are shown the results and comparisons between the thresholds obtained in the two methods. As shown, the thresholds obtained by the method presented in [3] shown a significant difference, in the thresholds, in the average gray levels and in the total points of each class. The result is more information included in the segmentation, which may make it more difficult for the recognition of regions of interest, both by specialists, as for systems to aid diagnosis (CAD). The increase of information in the image is displayed in Figure 7, where you can see that the method of Otsu identified parts of cardiac muscle as cell core or tissue rejection. TABLE IV COMPARISON OF THRESHOLDS (GRAY LEVELS) OBTAINED BY MAXIMUM ENTROPY AND BY OTSUS METHOD (1978), [3]. Otsu (1978) Maximum Entropy Difference Thresholds (Average) 139 203 70 240 26,95% -14,45%

Fig. 7 . Image (Fig. 3) processed with the Otsu's method [3]. The indication shows cardiac muscle included as the cell core or tissue rejection.

V. CONCLUSION According to used method, it was possible to find the images thresholds, and, therefore, segmenting them, presenting satisfactory results. Through comparisons between the techniques of thresholding by the lowest point and maximum entropy, we concluded that both techniques show good results, with irrelevant differences, and the second was defined as principal because it is a technique more widespread and accepted. Compared with the method shown in [3], for the studied images, the proposed technique shows better results because it allows the adjustment of parameters such as group size and slope percentage of the histogram, factors that influence the threshold values. Another important aspect is that the technique is based on the valleys analysis and in the characteristics of the histogram, and not only in probabilistic calculations. The method used in [3] includes false regions for the analysis of rejection, because in most cases, this method identifies higher values of thresholds, including not relevant information to the processing. This can lead to an erroneous diagnosis of the analyzed image. The proposed technique identify clearly cell core, fibrous tissue, muscle and tissue rejection, in myocardial

TABLE V COMPARISON OF AVERAGE GRAY LEVELS OBTAINED BY MAXIMUM ENTROPY AND BY OTSUS METHOD (1978), [3]. Average Gray Levels Class 1 Class 2 Class 3 104,91 172,44 232,38 41,12 165,99 243,55 24,92% 2,52% -4,36%
41

Otsu (1978) Maximum Entropy Difference

MARCH 15 - 19, 2010, LIMA, PERU

ISBN 978-1-4244-6294-0 IEEE CATALOG NUMBER CFP1018G-ART

PAN AMERICAN HEALTH CARE EXCHANGES PAHCE. CONFERENCE, WORKSHOPS, AND EXHIBITS. COOPERATION / LINKAGES. INTERCAMBIOS DE CUIDADO MDICO PANAMERICANOS. CONFERENCIA, TALLERES Y EXHIBICIONES. COOPERACIN / ENLACES.

images of biopsies from heart transplant patients, with advantages over one of the best known and widespread method in the literature. These characteristics are significant aspects of the developed technique, and allows the application to other image types, since the input parameters are adjustable to the studied case. This versatility and quality of results make the developed technique a considerable alternative to be applied during the stage of feature extraction in artificial vision systems.

REFERENCES
[1] A. K. Jain, R. Duin, Statistical pattern recognition: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22, no.1, pp. 437, 2000. [2] Gonzalez, R. C.; Woods, R. E. Processamento de imagens digitais. So Paulo: Edgard Blucher, 2000. [3] A. N. Otsu, A threshold selection method from gray-level histogram. IEEE Transactions on Systems, Man. and Cybernetics, pp. 62-66, 1978.

[4] T. A. Pun, New method for gray-level picture thresholding using the entropy of the histogram. Signal Processing, vol.2, pp. 223-237, 1980. [5] J. N. Kapur, P. K. Sahoo, K. C. Wong, A new method for graylevel picture thresholding using the entropy of the histogram. Computer Vision, Graphics, and Image Processing, vol.29, pp. 273-285, 1985. [6] A. H. Abutaleb, Automatic thresholding of gray-level pictures using two-dimensional entropy. Computer Vision, Graphics, and Image Processing, vol.47, pp. 22-32, 1989. [7] A. Beghdadi, A. L. Ngrate, P. V. De Lesegno, Entropic thresholding using a block source model. Computer Vision, Graphics, and Image Processing, vol.57, pp.197-205, 1995. [8] A. D. Brink, Using spatial information as an aid to maximum entropy image threshold selection. Pattern Recognition Letters, vol. 17, pp. 29-36, 1996. [9] S. R. Aboud Neta, L. V. Dutra, G. J. Erthal1, Limiarizao automtica em histogramas multimodais. Proceeedings of the 7th Brazilian Conference on Dynamics, Control and Applications, FCT Unesp de Presidente Prudente, May, 2008. [10] F. M. Lopes, Um modelo perceptivo de limiarizao de imagens digitais. Universidade Federal do Paran, 2003. Dissertao, Setor de Cincias Exatas Departamento de Informtica Mestrado em Informtica, Curitiba, 2003.

MARCH 15 - 19, 2010, LIMA, PERU

42

ISBN 978-1-4244-6294-0 IEEE CATALOG NUMBER CFP1018G-ART