Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

New Approach to Prostate Diagnosis - Perfusion CT Images Analysis using "Life Belt" Method

New Approach to Prostate Diagnosis - Perfusion CT Images Analysis using "Life Belt" Method The most important task that could improve the efficacy of managing the prostate cancer (PCa) is to develop the technique which will be able to detect an existing PCa even in cases when currently used methods are insufficient. It is supposed that the perfusion computed tomography technology (p-CT) can improve the diagnosis of early PCa. Unfortunately, the perfusion prostate images are very difficult to analyze especially for doctors who are not enough experienced with such a kind of images. Therefore there is a need to find a computational method which could help the doctors to make the decision whether the prostate cancer exists or not and (if the results are positive) to correctly point out the cancerous region. In research which results are presented in the paper we analyzed a great number of prostate images derived from over 50 patients with proven or suspected PCa. We propose the new method, named "life-belt" which has significant potential for identifying cancerous regions. KEYWORDS: prostate cancer, perfusion computed tomography, image processing, texture analysis Introduction The prostate cancer (PCa) is one of the most important medical problems. In many countries, for example in the USA it is the most popular men's cancer [1,24]. Although there are numerous methods and procedures for treating that malignancy, their successfulness strongly depends on the tumor progression. Only the PCa detected in early stage ­ before there arise metastasis ­ can be successful cured. Unfortunately it gives no readable symptoms and is very difficult for detection when is early enough. [5] Nowadays the most popular diagnostic procedures of PCa are the PSA protein measure and the DRA (per rectum) examination [7,29]. Both are suffered in too low level of sensitivity and specificity. The higher the PSA level the greater the likelihood of cancer. But the problem lies in the fact that the increase in the concentration of this protein is also observed in benign diseases, it is also a natural process associated with the patient's age. Hence the huge controversy such as setting norms for qualification a patient at risk [9,26,27,36,38]. The second method ­ the DRA study can capture only those changes that are perceptible in the peripheral zone of prostate apex. The only method which allows to confirm the existence of PCa is biopsy, at which a small portion of the gland is taken for histopathological examination. Of course, such a confirmation is possible only when the biopsy needle successfully hit into the pathologically changed part of the gland. Routinely accompanying biopsy transrectal ultrasound examination (TRUS) can help to indicate the suspected region, however it often does not work (when the changes are izoechogenic, invisible in TRUS). In such cases, a diagnostician is forced to collect tissue from randomly selected fragments of the gland, which is burdensome for the patient and may be fatal in consequences when the decision is incorrect. [8,25,32,37] Perfusion computed tomography For this reason, many researchers took the challenge to develop new imaging techniques, enabling increased diagnostic accuracy, especially in difficult cases. A team from the Cracow branch of the Oncology Center diagnoses patients with suspected PCa, by perfusion computed tomography (p-CT), collecting the necessary experimental material. There is a documented case where detection and location of the tumor was able thanks to p-CT, while biopsy under control of TRUS did not show anything. [19] The p-CT is a functional imaging technique. This method allows to evaluate the parameters of blood flow within diagnosed organs. It is documented in the literature that the growing tumor causes the creation of new blood vessels, which are irregular and characterized by increased permeability [4,6, 21,22]. It is supposed that the study of prostate perfusion will be able to identify those angiogenic areas, suspected to be with the presence of PCa. Currently, the p-CT imaging is successfully used mainly in the diagnosis of brain acute stroke, where the differences between healthy and ischemic area are much more easier to identify [15,17,23]. This method was also tested on other organs, including prostate [3,10,11,12,13,16,18,20,28,30,31,41]. However the published p-CT research in the field of prostate cancer diagnosis, were taken for a relatively small number of patients and were analyzed without any support from the advanced computational methods from image processing and pattern recognition area. p-CT prostate images are not clearly readable, so computational support can improve its value by detecting and evaluation of some features, difficult to see and assess. In addition the proposed technique can automatically analyze those features and point out the area suspected for PCa presence. Although Ives at all [18] claims that "Correlation between quantitative CT perfusion and tumor location is statistically significant only in subjects with localized high-volume, poorly differentiated prostate cancer.", we found an error in their reasoning [34]. We hope that the p-CT method may detect even early stage and low grade prostate cancer. Its advantage over TRUS is also that p-CT could show even isoechogenic cancer, invisible in TRUS guidance. Analyzed p-CT images were acquired using GE multislice scanner and created by its Advantage Workstation. It means that the images available to the authors are the result of not only an acquisition process, but also a postprocessing carried out by software associated to the imaging device and its console. Of course, it would be better to have access to the native data to improve the computation of the parameter maps or to perform the detection of cancerous tissues directly from them, without explicit use of the parameter maps. However, the authors did not have such possibilities and had to use already preprocessed parameter maps on which the prostate shape was deformed into ellipse (Figure 1). Figure 1. Example of p-CT image ­ the minor pelvis cross-section map. The oval marks the area of the prostate. During the p-CT examination the four parameters are measured: blood flow (BF) ­ the volume of blood flowing per unit time within a given area; blood volume (BV) ­ the total volume of blood in the diagnosed area; mean transit time (MTT) ­ specified in seconds, average blood flow through the tissue; permeability surface (PS) ­ measures the vascular permeability of bolus migrating from intravascular to extravascular space. Examples of different parameters are presented on Figure 2. On those images only the prostate area presented. The prostate selection from the whole minor pelvis cross-section was performed manually keeping an aforementioned elliptic shape. Figure 2. The p-CT prostate images of an exemplary patient: a) blood flow (BF); b) blood volume (BV); c) mean transit time (MTT); d) permeability surface (PS). Prostate texture analysis Parametric maps (BF, BV, MTT and PS) were drawn at three levels (conventionally base, middle and apex) of the gland. In order to perform the computational analysis, only the area of prostate was selected from the acquired images. The images, originally coded with pseudocolor, where blue symbolizes the area with minimal, and red ­ the area with maximal perfusion, were transformed into the grayscale using a linear transformation without loose of any information. To ensure the best quality of images some image preprocessing methods were used like scaling and histogram equalization. After transformation the size of each analyzed image was 100x120 pixels. For automatic description of the texture of particular regions on analyzed p-CT images the gray level co-occurrence matrices (GLCM) [14] were calculated. Let I : Z2D G = { 1, ..., Ng } (where Z denotes set of integers) be a two-dimensional discrete image with Ng gray levels. For the given image I we define the GLCM: P(i, j | d , ) #{k , l D : I (k ) i, I (l ) j, || k l || d , (k l ) } , #{m, n D : || m n || d , (m n) } (1) where: i,j G ­ gray levels of points k and l, respectively; (k l ) - the angle between vector kl and axe 0 X ; d - distance between k and l; - direction of co-occurrence, #X ­ power (number of elements) of set X. GLCM allow us to evaluate a number of coefficients, which characterize the texture of the analyzed image. In our research 21 different coefficients [33] were calculated for each matrix characterized by distance d in range 1 to 9, and angle with values 0° and 90°. So for each analyzed perfusion parameter we obtained 378-dimensional feature space. Resulted values for each feature were analyzed in order to eliminate outliers and normalized. The distribution of each feature was equalized using the ladder of powers method [39, 40] with (0,2]. In this method the function error() is defined: error ( ) : [cdf{x } { x , var (x )}] 2 , c c c c 1,2 x (2) where: c = {1,2} ­ classification; cdf ( xc ), xc , var ( xc ) ­ distribution function, mean and variance of empirical distribution for class c, respectively; (,2) ­ normal distribution function with mean and variance 2. We were looking for opt, which minimize error(): opt min {error ( )} . (3) Features, where error(opt)1 were excluded from further analysis. Our goal was to select such a subspace that consist of small number of features which are not correlated and used together have the best discriminatory power. The discriminatory power was measured for each analyzed feature subspace using the Bhattacharyya measure for normal distribution [2]: Jx | 1 (1 2 ) | 1 1 ( 1 2 )T 1 2 1 ( 1 2 ) log 2 4 2 | 1 || 2 | (4) where: 1, 2 ­ means, 1, 2 ­ covariance matrices of x feature subspace for classes 1 and 2, respectively, || ­ determinant . The algorithm was described in detail and discussed in our another work. [35] Region of interest For further research work described in this paper, every analyzed image has been divided into smaller fragments (regions of interest ­ ROI). Thus, for each considered ROI were determined and verified the parameters characterizing the texture of the area in order to find local irregularities. There are many theoretically possible ways to determine the size and shape of ROI. Some of them were discussed in our earlier work [33]. We tested different rectangular masks (as well square as those with longer vertical or horizontal edge) and also oval ROIs. For main experiments with large set of images following two strategies were used: 1. 2. coverage area of a rectangular mask; the "life belt" method. In the first attempt, the entire image is covered by a mask with given size and shape. The previous considerations lead the authors to select a rectangular mask. Moreover, due to the visibility of anisotropy, the mask should have a vertical direction. The mask size was set to 10x20 pixels. The idea of the "life-belt" method is based on the following facts: - Perfusion changes in the middle of the image (roughly corresponding to a transition zone and a central zone of the prostate) are usually connected with benign prostatic hyperplasia (BPH) (Figure 3). Therefore, this area was isolated and excluded from further consideration. Figure 3. Increased perfusion level in the central part of the picture is usually caused by benign prostatic hyperplasia (BPH): a) the example of prostate with BPH; b) the same image with the "life-belt" mask - the central area was excluded from the analysis. Cancer usually appears in the prostate peripheral zone. However, zonal structure is not visible in the p-CT examination. Therefore it is rather impossible to precisely select the peripheral zone. In addition, the zonal structure differs from base to apex of prostate. It follows that we should accept some simplifications. Thus proposed ROI shape is the same for all images and not exactly consider true anatomy. It is only approximation. In some difficult cases, when the pathological changes are invisible on TRUS, they can be find thanks to the asymmetry of the outline of the prostate. It cannot be a rule, because the cancer sometimes takes the multifocal form, present in both lobes. Nevertheless, a healthy prostate is symmetrical, therefore the symmetry analysis may be a source of extremely valuable information. For further analysis the central elliptic area of size 30 (horizontal) x 25 (vertical) pixels was excluded. The above number are the result of aforementioned agreement (the ROI should be the same for each image, we also remember that each of our prostate images are the same in size) and the images analysis which shown that those values would perform the best. The remaining area (which generally is also elliptic) was divided into 6 parts - three for each lobe of the prostate. There were also variations of this pattern, consisting of smaller fragments. Finally, for the experiments described below three patterns of ROI selection were used. In each of these the central area of the gland was excluded from considerations: 1. The image was covered by a rectangular masks, sized 10x20 pixels each. Masks were determined with the jump of 10 pixels. The corners and the central part of the image were omitted. For each image 76 areas of ROI were selected (Figure 4). Figure 4. Rectangular mask. a, b) the location of each ROI (each ROI is presented with a different shade; due to the partially overlap those areas are shown on the two images, c) an image coverage by ROIs. Large "life-belt" ­ each ROI includes the fragment of the ellipse (without its central part) with a width of 60°. Following areas are determined by turn 20 degrees, so also here we have a partial overlapping of neighboring ROIs. For each image 18 ROI areas were selected (Figure 5). Figure 5. Large "life belt" - due to the partial overlapping of the different areas, following examples of the ROIs are shown in separate images. The same color identify symmetrical pairs. Small "life-belt" ­ a variant of the previous version, in which ROIs are smaller - each has a width of 20°. The areas are disjoint, and there are also 18 ROIs for each image (Figure 6). Figure 6. Small "life-belt". Symmetrical pairs are determined by the same color. Symmetric analysis The analysis of symmetrical differences within the prostate may result with valuable information. Thus the calculation was made in two ways: standard analysis, where each ROI was considered individually, and "symmetric analysis". The second approach is based on the above mentioned assumption that the symmetrical differences within corresponding prostate lobes may suggest the presence of cancer. On figures 4 and 5 matching pairs of areas are pointed over the same color. The first step was the same as in standard analysis ­ for each individual ROI the values of all features were evaluated. Then for each corresponding pair the difference between left and right ROI were calculated and normalized to the [0,1] range. So the value 0,5 means no differences between left and right area; the value close to 0 means the domination of the left lobe; and ­ similarly ­ the value close to 1 means the domination of the right lobe. It should be noted that the pairs of ROI, where the same area must be calculated twice, were omitted. Thus in the "large life-belt" only 7 pairs was taken under consideration (Table 1). Table 1. The number of analyzed ROIs within an image. analysis rectangular mask standard 76 symmetrical 38 large ,,life-belt" 18 7 small ,,life-belt" 18 9 Anisotropy measure Apart of analysis of cancerous prostates, in our work also one healthy (without PCa) patient was examined. It was observed on his p-CT image that the texture of healthy area tends to be directional in opposite to the texture of the cancerous area (Figure 7). Therefore we prepared some tests used to measure the degree of anisotropy. The applied algorithm was very similar to other presented in this paper. The values of features were based on the differences between the corresponding features calculated for GLCMs with given displacement d but perpendicular angles. Figure 7. The comparison of images for healthy (a) and cancerous (b) patient. In healthy areas the regions with increased perfusion are rather horizontal, while in cancerous region (left side of image b) not. Results The total number of 240 experiments was conducted. In each test we used images from one of the automatically (during p-CT examination) measured perfusion parameters (BF, BV, MTT or PS) or created new images (the parameter "ALL") by taking into consideration source images for all of the above mentioned parameters simultaneously: ALL = BF + BV ­ MTT + PS. The experiments differed also with preprocessing methods (eg. scaling techniques, histogram or reference values equalization) and ROI selection techniques (we used 3 of the described above attempts). For each possible parameters combination we tested both standard and "symmetric" analysis method and finally isotropic and anisotropic attempt. In each experiment we identified the potentially best set of six features (with the highest value of the Bhattacharya measure). In order to select the cancerous area we used the quadratic decision function. For each calculation we analyzed and tested 59 p-CT images. As the best occurs the analysis of images based on all perfusion parameters, with standard isotropic analysis, and the "small life-belt" method of ROI selection. The preprocessing methods were scaling images into the size 120x100 using the nearest neighbor method and histogram equalization. In this experiment we achieved about 86% of correct identifications of the cancerous area. Figure 8 presents the distribution of features in the mentioned above experiment. While it is difficult to present the 6-dimensional space, the discriminatory power for each pair is presented separately. Figure 8. Features for the best recognition. Green ­ healthy areas; red ­ cancerous areas. In consecutive subimages the values for pairs of features are presented, respectively: a) 1-2; b) 1-3; c) 1-4; d) 1-5; e) 1-6; f) 2-3; g) 2-4; h) 2-5; i) 2-6; j) 3-4; k) 3-5; l) 3-6; m) 4-5; n) 4-6; o) 5-6, where 1: d=1, =90°, f9; 2: d=2, =0°, f6; 3: d=4, =90°, f3; 4: d=4, =90°, f6; 5: d=5, =90°, f11; 6: d=6, =90°, f1. Conclusion It seems that the perfusion computer tomography (p-CT) may be helpful in detecting the cancerous lesions within prostate, and therefore can significantly improve the diagnosis of the early PCa. This method is especially advised for cases which are too hard for traditional diagnostic methods, like for example TRUS. Although the p-CT prostate images are very difficult for analysis, the presented computational methods are very helpful for diagnosticians and makes such an analysis much easier. Especially important in this paper is the presentation of the "life-belt" method of ROI selection. This original technique occurs the best of all methods we tested. All of the measured perfusion parameters (BF, BV, MTT, PS) are important and each of them has a valuable impact to the quality of recognition. Our other ideas - the symmetric analysis and the anisotropy measure did not occur as good as we supposed. Of course, the 86% of correct recognition is not enough and the attempts to improve the proposed algorithm should be the subject of further analysis. However, it must be noted that our result is much better that in the previous publication concerning this problem, where the advanced computational methods were not used [28]. It is one more proof that the computational analysis is an essential in the p-CT prostate images analysis. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Bio-Algorithms and Med-Systems de Gruyter

New Approach to Prostate Diagnosis - Perfusion CT Images Analysis using "Life Belt" Method

Download PDF
14 pages

Loading next page...
 
/lp/de-gruyter/new-approach-to-prostate-diagnosis-perfusion-ct-images-analysis-using-o9oob0fXN0
Publisher
de Gruyter
Copyright
Copyright © 2012 by the
ISSN
1895-9091
eISSN
1896-530X
DOI
10.2478/bams-2012-0009
Publisher site
See Article on Publisher Site

Abstract

The most important task that could improve the efficacy of managing the prostate cancer (PCa) is to develop the technique which will be able to detect an existing PCa even in cases when currently used methods are insufficient. It is supposed that the perfusion computed tomography technology (p-CT) can improve the diagnosis of early PCa. Unfortunately, the perfusion prostate images are very difficult to analyze especially for doctors who are not enough experienced with such a kind of images. Therefore there is a need to find a computational method which could help the doctors to make the decision whether the prostate cancer exists or not and (if the results are positive) to correctly point out the cancerous region. In research which results are presented in the paper we analyzed a great number of prostate images derived from over 50 patients with proven or suspected PCa. We propose the new method, named "life-belt" which has significant potential for identifying cancerous regions. KEYWORDS: prostate cancer, perfusion computed tomography, image processing, texture analysis Introduction The prostate cancer (PCa) is one of the most important medical problems. In many countries, for example in the USA it is the most popular men's cancer [1,24]. Although there are numerous methods and procedures for treating that malignancy, their successfulness strongly depends on the tumor progression. Only the PCa detected in early stage ­ before there arise metastasis ­ can be successful cured. Unfortunately it gives no readable symptoms and is very difficult for detection when is early enough. [5] Nowadays the most popular diagnostic procedures of PCa are the PSA protein measure and the DRA (per rectum) examination [7,29]. Both are suffered in too low level of sensitivity and specificity. The higher the PSA level the greater the likelihood of cancer. But the problem lies in the fact that the increase in the concentration of this protein is also observed in benign diseases, it is also a natural process associated with the patient's age. Hence the huge controversy such as setting norms for qualification a patient at risk [9,26,27,36,38]. The second method ­ the DRA study can capture only those changes that are perceptible in the peripheral zone of prostate apex. The only method which allows to confirm the existence of PCa is biopsy, at which a small portion of the gland is taken for histopathological examination. Of course, such a confirmation is possible only when the biopsy needle successfully hit into the pathologically changed part of the gland. Routinely accompanying biopsy transrectal ultrasound examination (TRUS) can help to indicate the suspected region, however it often does not work (when the changes are izoechogenic, invisible in TRUS). In such cases, a diagnostician is forced to collect tissue from randomly selected fragments of the gland, which is burdensome for the patient and may be fatal in consequences when the decision is incorrect. [8,25,32,37] Perfusion computed tomography For this reason, many researchers took the challenge to develop new imaging techniques, enabling increased diagnostic accuracy, especially in difficult cases. A team from the Cracow branch of the Oncology Center diagnoses patients with suspected PCa, by perfusion computed tomography (p-CT), collecting the necessary experimental material. There is a documented case where detection and location of the tumor was able thanks to p-CT, while biopsy under control of TRUS did not show anything. [19] The p-CT is a functional imaging technique. This method allows to evaluate the parameters of blood flow within diagnosed organs. It is documented in the literature that the growing tumor causes the creation of new blood vessels, which are irregular and characterized by increased permeability [4,6, 21,22]. It is supposed that the study of prostate perfusion will be able to identify those angiogenic areas, suspected to be with the presence of PCa. Currently, the p-CT imaging is successfully used mainly in the diagnosis of brain acute stroke, where the differences between healthy and ischemic area are much more easier to identify [15,17,23]. This method was also tested on other organs, including prostate [3,10,11,12,13,16,18,20,28,30,31,41]. However the published p-CT research in the field of prostate cancer diagnosis, were taken for a relatively small number of patients and were analyzed without any support from the advanced computational methods from image processing and pattern recognition area. p-CT prostate images are not clearly readable, so computational support can improve its value by detecting and evaluation of some features, difficult to see and assess. In addition the proposed technique can automatically analyze those features and point out the area suspected for PCa presence. Although Ives at all [18] claims that "Correlation between quantitative CT perfusion and tumor location is statistically significant only in subjects with localized high-volume, poorly differentiated prostate cancer.", we found an error in their reasoning [34]. We hope that the p-CT method may detect even early stage and low grade prostate cancer. Its advantage over TRUS is also that p-CT could show even isoechogenic cancer, invisible in TRUS guidance. Analyzed p-CT images were acquired using GE multislice scanner and created by its Advantage Workstation. It means that the images available to the authors are the result of not only an acquisition process, but also a postprocessing carried out by software associated to the imaging device and its console. Of course, it would be better to have access to the native data to improve the computation of the parameter maps or to perform the detection of cancerous tissues directly from them, without explicit use of the parameter maps. However, the authors did not have such possibilities and had to use already preprocessed parameter maps on which the prostate shape was deformed into ellipse (Figure 1). Figure 1. Example of p-CT image ­ the minor pelvis cross-section map. The oval marks the area of the prostate. During the p-CT examination the four parameters are measured: blood flow (BF) ­ the volume of blood flowing per unit time within a given area; blood volume (BV) ­ the total volume of blood in the diagnosed area; mean transit time (MTT) ­ specified in seconds, average blood flow through the tissue; permeability surface (PS) ­ measures the vascular permeability of bolus migrating from intravascular to extravascular space. Examples of different parameters are presented on Figure 2. On those images only the prostate area presented. The prostate selection from the whole minor pelvis cross-section was performed manually keeping an aforementioned elliptic shape. Figure 2. The p-CT prostate images of an exemplary patient: a) blood flow (BF); b) blood volume (BV); c) mean transit time (MTT); d) permeability surface (PS). Prostate texture analysis Parametric maps (BF, BV, MTT and PS) were drawn at three levels (conventionally base, middle and apex) of the gland. In order to perform the computational analysis, only the area of prostate was selected from the acquired images. The images, originally coded with pseudocolor, where blue symbolizes the area with minimal, and red ­ the area with maximal perfusion, were transformed into the grayscale using a linear transformation without loose of any information. To ensure the best quality of images some image preprocessing methods were used like scaling and histogram equalization. After transformation the size of each analyzed image was 100x120 pixels. For automatic description of the texture of particular regions on analyzed p-CT images the gray level co-occurrence matrices (GLCM) [14] were calculated. Let I : Z2D G = { 1, ..., Ng } (where Z denotes set of integers) be a two-dimensional discrete image with Ng gray levels. For the given image I we define the GLCM: P(i, j | d , ) #{k , l D : I (k ) i, I (l ) j, || k l || d , (k l ) } , #{m, n D : || m n || d , (m n) } (1) where: i,j G ­ gray levels of points k and l, respectively; (k l ) - the angle between vector kl and axe 0 X ; d - distance between k and l; - direction of co-occurrence, #X ­ power (number of elements) of set X. GLCM allow us to evaluate a number of coefficients, which characterize the texture of the analyzed image. In our research 21 different coefficients [33] were calculated for each matrix characterized by distance d in range 1 to 9, and angle with values 0° and 90°. So for each analyzed perfusion parameter we obtained 378-dimensional feature space. Resulted values for each feature were analyzed in order to eliminate outliers and normalized. The distribution of each feature was equalized using the ladder of powers method [39, 40] with (0,2]. In this method the function error() is defined: error ( ) : [cdf{x } { x , var (x )}] 2 , c c c c 1,2 x (2) where: c = {1,2} ­ classification; cdf ( xc ), xc , var ( xc ) ­ distribution function, mean and variance of empirical distribution for class c, respectively; (,2) ­ normal distribution function with mean and variance 2. We were looking for opt, which minimize error(): opt min {error ( )} . (3) Features, where error(opt)1 were excluded from further analysis. Our goal was to select such a subspace that consist of small number of features which are not correlated and used together have the best discriminatory power. The discriminatory power was measured for each analyzed feature subspace using the Bhattacharyya measure for normal distribution [2]: Jx | 1 (1 2 ) | 1 1 ( 1 2 )T 1 2 1 ( 1 2 ) log 2 4 2 | 1 || 2 | (4) where: 1, 2 ­ means, 1, 2 ­ covariance matrices of x feature subspace for classes 1 and 2, respectively, || ­ determinant . The algorithm was described in detail and discussed in our another work. [35] Region of interest For further research work described in this paper, every analyzed image has been divided into smaller fragments (regions of interest ­ ROI). Thus, for each considered ROI were determined and verified the parameters characterizing the texture of the area in order to find local irregularities. There are many theoretically possible ways to determine the size and shape of ROI. Some of them were discussed in our earlier work [33]. We tested different rectangular masks (as well square as those with longer vertical or horizontal edge) and also oval ROIs. For main experiments with large set of images following two strategies were used: 1. 2. coverage area of a rectangular mask; the "life belt" method. In the first attempt, the entire image is covered by a mask with given size and shape. The previous considerations lead the authors to select a rectangular mask. Moreover, due to the visibility of anisotropy, the mask should have a vertical direction. The mask size was set to 10x20 pixels. The idea of the "life-belt" method is based on the following facts: - Perfusion changes in the middle of the image (roughly corresponding to a transition zone and a central zone of the prostate) are usually connected with benign prostatic hyperplasia (BPH) (Figure 3). Therefore, this area was isolated and excluded from further consideration. Figure 3. Increased perfusion level in the central part of the picture is usually caused by benign prostatic hyperplasia (BPH): a) the example of prostate with BPH; b) the same image with the "life-belt" mask - the central area was excluded from the analysis. Cancer usually appears in the prostate peripheral zone. However, zonal structure is not visible in the p-CT examination. Therefore it is rather impossible to precisely select the peripheral zone. In addition, the zonal structure differs from base to apex of prostate. It follows that we should accept some simplifications. Thus proposed ROI shape is the same for all images and not exactly consider true anatomy. It is only approximation. In some difficult cases, when the pathological changes are invisible on TRUS, they can be find thanks to the asymmetry of the outline of the prostate. It cannot be a rule, because the cancer sometimes takes the multifocal form, present in both lobes. Nevertheless, a healthy prostate is symmetrical, therefore the symmetry analysis may be a source of extremely valuable information. For further analysis the central elliptic area of size 30 (horizontal) x 25 (vertical) pixels was excluded. The above number are the result of aforementioned agreement (the ROI should be the same for each image, we also remember that each of our prostate images are the same in size) and the images analysis which shown that those values would perform the best. The remaining area (which generally is also elliptic) was divided into 6 parts - three for each lobe of the prostate. There were also variations of this pattern, consisting of smaller fragments. Finally, for the experiments described below three patterns of ROI selection were used. In each of these the central area of the gland was excluded from considerations: 1. The image was covered by a rectangular masks, sized 10x20 pixels each. Masks were determined with the jump of 10 pixels. The corners and the central part of the image were omitted. For each image 76 areas of ROI were selected (Figure 4). Figure 4. Rectangular mask. a, b) the location of each ROI (each ROI is presented with a different shade; due to the partially overlap those areas are shown on the two images, c) an image coverage by ROIs. Large "life-belt" ­ each ROI includes the fragment of the ellipse (without its central part) with a width of 60°. Following areas are determined by turn 20 degrees, so also here we have a partial overlapping of neighboring ROIs. For each image 18 ROI areas were selected (Figure 5). Figure 5. Large "life belt" - due to the partial overlapping of the different areas, following examples of the ROIs are shown in separate images. The same color identify symmetrical pairs. Small "life-belt" ­ a variant of the previous version, in which ROIs are smaller - each has a width of 20°. The areas are disjoint, and there are also 18 ROIs for each image (Figure 6). Figure 6. Small "life-belt". Symmetrical pairs are determined by the same color. Symmetric analysis The analysis of symmetrical differences within the prostate may result with valuable information. Thus the calculation was made in two ways: standard analysis, where each ROI was considered individually, and "symmetric analysis". The second approach is based on the above mentioned assumption that the symmetrical differences within corresponding prostate lobes may suggest the presence of cancer. On figures 4 and 5 matching pairs of areas are pointed over the same color. The first step was the same as in standard analysis ­ for each individual ROI the values of all features were evaluated. Then for each corresponding pair the difference between left and right ROI were calculated and normalized to the [0,1] range. So the value 0,5 means no differences between left and right area; the value close to 0 means the domination of the left lobe; and ­ similarly ­ the value close to 1 means the domination of the right lobe. It should be noted that the pairs of ROI, where the same area must be calculated twice, were omitted. Thus in the "large life-belt" only 7 pairs was taken under consideration (Table 1). Table 1. The number of analyzed ROIs within an image. analysis rectangular mask standard 76 symmetrical 38 large ,,life-belt" 18 7 small ,,life-belt" 18 9 Anisotropy measure Apart of analysis of cancerous prostates, in our work also one healthy (without PCa) patient was examined. It was observed on his p-CT image that the texture of healthy area tends to be directional in opposite to the texture of the cancerous area (Figure 7). Therefore we prepared some tests used to measure the degree of anisotropy. The applied algorithm was very similar to other presented in this paper. The values of features were based on the differences between the corresponding features calculated for GLCMs with given displacement d but perpendicular angles. Figure 7. The comparison of images for healthy (a) and cancerous (b) patient. In healthy areas the regions with increased perfusion are rather horizontal, while in cancerous region (left side of image b) not. Results The total number of 240 experiments was conducted. In each test we used images from one of the automatically (during p-CT examination) measured perfusion parameters (BF, BV, MTT or PS) or created new images (the parameter "ALL") by taking into consideration source images for all of the above mentioned parameters simultaneously: ALL = BF + BV ­ MTT + PS. The experiments differed also with preprocessing methods (eg. scaling techniques, histogram or reference values equalization) and ROI selection techniques (we used 3 of the described above attempts). For each possible parameters combination we tested both standard and "symmetric" analysis method and finally isotropic and anisotropic attempt. In each experiment we identified the potentially best set of six features (with the highest value of the Bhattacharya measure). In order to select the cancerous area we used the quadratic decision function. For each calculation we analyzed and tested 59 p-CT images. As the best occurs the analysis of images based on all perfusion parameters, with standard isotropic analysis, and the "small life-belt" method of ROI selection. The preprocessing methods were scaling images into the size 120x100 using the nearest neighbor method and histogram equalization. In this experiment we achieved about 86% of correct identifications of the cancerous area. Figure 8 presents the distribution of features in the mentioned above experiment. While it is difficult to present the 6-dimensional space, the discriminatory power for each pair is presented separately. Figure 8. Features for the best recognition. Green ­ healthy areas; red ­ cancerous areas. In consecutive subimages the values for pairs of features are presented, respectively: a) 1-2; b) 1-3; c) 1-4; d) 1-5; e) 1-6; f) 2-3; g) 2-4; h) 2-5; i) 2-6; j) 3-4; k) 3-5; l) 3-6; m) 4-5; n) 4-6; o) 5-6, where 1: d=1, =90°, f9; 2: d=2, =0°, f6; 3: d=4, =90°, f3; 4: d=4, =90°, f6; 5: d=5, =90°, f11; 6: d=6, =90°, f1. Conclusion It seems that the perfusion computer tomography (p-CT) may be helpful in detecting the cancerous lesions within prostate, and therefore can significantly improve the diagnosis of the early PCa. This method is especially advised for cases which are too hard for traditional diagnostic methods, like for example TRUS. Although the p-CT prostate images are very difficult for analysis, the presented computational methods are very helpful for diagnosticians and makes such an analysis much easier. Especially important in this paper is the presentation of the "life-belt" method of ROI selection. This original technique occurs the best of all methods we tested. All of the measured perfusion parameters (BF, BV, MTT, PS) are important and each of them has a valuable impact to the quality of recognition. Our other ideas - the symmetric analysis and the anisotropy measure did not occur as good as we supposed. Of course, the 86% of correct recognition is not enough and the attempts to improve the proposed algorithm should be the subject of further analysis. However, it must be noted that our result is much better that in the previous publication concerning this problem, where the advanced computational methods were not used [28]. It is one more proof that the computational analysis is an essential in the p-CT prostate images analysis.

Journal

Bio-Algorithms and Med-Systems – de Gruyter

Published: Jan 1, 2012

There are no references for this article.