Non-invasive evaluation of hepatic steatosis by ... - Semantic Scholar

2 downloads 0 Views 212KB Size Report
Sep 27, 2007 - Abstract — The gold standard for the quantitative evaluation of steatosis is liver biopsy, but this is an invasive method. The recent trend is to ...
1st International Conference on Advancements of Medicine and Health Care through Technology, MediTech2007, 27-29th September, 2007, Cluj-Napoca, ROMANIA

Non-invasive evaluation of hepatic steatosis by ultrasound image analysis with simple brightness features and support vector machines Georgiana Nagy, Mihaela Gordan, A. Vlaicu, P. A. Mircea, Doiniţa Crişan, and Simona Vălean Abstract — The gold standard for the quantitative evaluation of steatosis is liver biopsy, but this is an invasive method. The recent trend is to investigate and develop novel non-invasive hepatic tissue evaluation methods, able to give performances close to biopsy, one of the possible solutions being ultrasound image analysis. Finding the most suitable descriptors of the histological tissue changes (invariant to the ultrasound device and patient) in this imaging modality is an important issue. Several such descriptors are reported in the literature, starting from simple intensity parameters to more sophisticated ones. Here we investigate the discrimination ability of the hepatic steatosis as opposed to healthy tissue by a set of computationally simple features, extracted from the gray level histogram of a small sized region of interest (ROI) positioned at three depths in the ultrasound hepatic image. The advantage of finding computationally simple features, as intensity histogram extracted features, can be the possibility to include their computation directly in ultrasound devices, provided they offer good tissue discrimination. The in-depth variations of some of the features investigated in this paper show a good discrimination in steatosis evaluation and quantification, close to the state of the art in the field. Keywords: Steatosis, non-invasive tissue analysis, ultrasound imaging, grey level histogram features, tissue classification, support vector machines.

Hepatic steatosis is an anatomo-clinical entity, considered as a benign disease with limited evolution potential, but it was recently proven that simple steatosis can evolve towards hepatic fibrosis. Having knowledge of the evolution stages in the hepatic tissue modification from steatosis to steato-hepatitis, to fibrosis and cirrhosis is essential for diagnosis and for understanding the pathogenesis of disease progression, with immediate consequences regarding therapy.

1. INTRODUCTION One of the most actual topics in the hepatological research is the investigation of chronic hepatic diseases (steatosis, non-alcoholic steatohepatitis, viral and autoimmune chronic hepatitis). The reasons for this increased interest are the incomplete knowledge about the pathogenesis and the factors influencing the natural evolution of these diseases, on one hand, and the necessity to identify prophylactic and therapeutic solutions, on the other hand. Chronic hepatic diseases constitute an important public health problem, by increased incidence, difficulties in prophylaxis and therapy and especially by the evolution in many cases towards cirrhosis and hepatocarcinoma [1], influencing the survival and life quality [2].

The diffuse hepatic steatosis is known to be visually noticeable in ultrasonography by the pathological increase of the tissue brightness and of the posterior attenuation [3, 4]. In general, the ultrasound examination is performed visually by the physician, which makes the ultrasound image interpretation is subjective (operator dependent), not by quantitative evaluation. This approach makes the use of ultrasonography limited as non-invasive diagnosis method, especially making practically impossible a reliable time monitoring of hepatic steatosis evolution [5, 6, 7]. This motivates the current trend in the computer ultrasound imaging community, of developing computer-based algorithms to extract suitable quantifiable descriptors of steatosis from ultrasound hepatic images, able to accurately discriminate the hepatic tissue changes. Once a set of such descriptors is available, they can either be interpreted manually or automatically using classification systems that can discriminate between healthy tissue and steatosis. The final goal is to obtain computer-based systems to aid the physician in the evaluation of the hepatic tissue modifications and monitoring of their evolution based on ultrasound imaging alone, while providing performances close to the ones obtained by the classical biopsy

G. Nagy is with the University of Medicine and Pharmacy of ClujNapoca, Romania, phone: +40-264-406-839; e-mail: georgi_nagy @ yahoo.com. M. Gordan is with the Technical University of Cluj-Napoca, Romania, phone: +40-264-401-309; e-mail: Mihaela.Gordan @ bel.utcluj.ro. A. Vlaicu is with the Technical University of Cluj-Napoca, Romania, phone: +40-264-401-204; e-mail: Aurel.Vlaicu @ bel.utcluj.ro. P.A. Mircea is with the University of Medicine and Pharmacy of Cluj-Napoca, Romania, phone: +40-264-406-839; e-mail: pmircea @ umfcluj.ro. D. Crisan is with the University of Medicine and Pharmacy of ClujNapoca, Romania, phone: +40-264-406-839; e-mail: doinitacrisan @ gmail.com. S. Valean is with the University of Medicine and Pharmacy of ClujNapoca, Romania, phone: +40-264-406-839; e-mail: simonavalean @ hotmail.com.

217

1st International Conference on Advancements of Medicine and Health Care through Technology, MediTech2007, 27-29th September, 2007, Cluj-Napoca, ROMANIA comparing the brightness of the steatosis with the healthy tissue, we consider as an extra-descriptor of the local brightness, the median brightness of the patient’s liver referenced to the median brightness of the same patient’s spleen extracted from the corresponding ultrasound image acquired in the same examination session.

approach [8, 9, 10]. The extracted descriptors should be invariant to the type of ultrasound device and patient’s physiological variability. Several such hepatic tissue descriptors extracted from ultrasound images are reported in the literature, of variable computational complexity, from simple intensity features to more sophisticated ones. Some of them, as texture co-occurrence features or fractal dimension, prove good ability to discriminate hepatic modifications [9], but are computational expensive. The advantage of finding computationally simple descriptors (as e.g. the intensity histogram extracted features), at similar discrimination performance to more complex ones, would be the possibility to include their computation directly in ultrasound devices, since some of these devices already have incorporated the ability of computing the histogram of a selected ROI, and based on it, simple parameters as median luminance, standard deviation. Thus embedding in the ultrasound machine’s microprocessor the computation of other simple histogram based parameters, as the ones used in this paper, is a feasible task. This is one of the motivations for investigating here the discriminative power of such simple histogram descriptors in histological tissue changes evaluation and quantification. Unlike most existing methods [8, 10, 11], we believe that the descriptors discriminative accuracy can be increased if we consider not only one single descriptor value per image, but also its in-depth variation, as suggested by the human fashion of ultrasound image investigation for steatosis identification. These particular descriptors, their use in an automatic classification system and their results in the identification of steatosis are described in the following.

The concept of “local neighbourhood” of the ultrasound image, where the brightness distribution is evaluated and quantified with the extracted features, is defined again making use of medical expert’s opinion: a square region of interest of moderate size is chosen to describe locally the “perceived” brightness of the tissue patch. The indepth brightness variation is obtained by computing these features in regions of interests positioned at three consecutive fixed depths below the hepatic capsule. Once the numerical features describing the brightness at the three selected depths are computed for a hepatic ultrasound image, we should assess their ability to discriminate steatosis from healthy hepatic tissue. This task can be formulated as a binary classification problem, which can be implemented using e.g. a support vector machine (SVM) classifier (this is the solution we adopt). The classifier labels each ultrasound image as presenting steatosis or not. The classification accuracy (i.e. the match between the classification results and the real known diagnostic, provided by the biopsy) is a good indicator of the suitability of each examined feature in the discrimination of severe steatosis. This system was implemented in the form of a Windows application. The user can load the spleen or the hepatic ultrasound images and can position the ROI at specified depths. The application processes the brightness contents in the ROI to extract the parameter values and saves these values in text files to be further used in the SVM classification training and SVM classifier test phase. The interface of this application and an example window of extracted features is presented in Figure 1.

2. THE PROPOSED ULTRASOUND IMAGE ANALYSIS METHOD FOR HEPATIC STEATOSIS EVALUATION

The method we propose aims to implement in a quantitative (objective) fashion the qualitative formulated medical expert knowledge concerning the correlation between the in-depth mean brightness variation in the ultrasound image and the presence of steatosis in the hepatic tissue – variation that should not appear in the healthy liver tissue [3, 4]. However this is a non-trivial task, especially in what concerns the choice of the quantifiable features to be used for analysing the in-depth brightness variation, since it is not obvious whether when examining visually the hepatic ultrasound image, the human visual system actually perceives its “average” brightness as the mean or median luminance in a limited sized neighbourhood or on the contrary the “weights” of the different grey levels is different in the perception. The second situation is many times likely to occur. Therefore although the most straightforward parameter to describe the in-depth intensity attenuation in the ultrasound image is the median of the brightness, other simple statistical grey level distribution parameters can also be important for the quantitative description of the in-depth attenuation, as e.g. the darkest and the brightest grey level in the neighbourhood, spread and skew of the histogram. Furthermore to minimize the inter-subject variability in

2.1. ROI Selection and Feature Extraction According to the examination procedure of the ultrasound images described above, the description of the brightness variation at progressive depths in the hepatic ultrasound image is done not at pixel level, but based on a local characterisation of the "average" brightness. The definition of "local" usually varies depending on the image resolution and hepatic ultrasound image size. In our case, for the study group acquired and used in our experiments, the physician examination and experience suggested as local neighbourhood used as ROI for the grey level distribution description, a small square region of size 1cm×1cm. This corresponds, in pixels, to 40×40 pixels in each ROI, for the current acquisition conditions. Since the mean depth of the hepatic images acquired is 7 cm - 8 cm, we consider as suitable depths for the examination of the "average" brightness variation the values: 0.5 cm, 2 cm and 3.5 cm in depth referenced to the hepatic capsule (considered as 0 depth). Thus, to 218

1st International Conference on Advancements of Medicine and Health Care through Technology, MediTech2007, 27-29th September, 2007, Cluj-Napoca, ROMANIA extract the quantitative statistical descriptors for the local brightness in each hepatic ultrasound image, the ROI of fixed size as defined here is successively placed at the three depths. In each position, several features are computed from the histogram of the grey levels in the ROI, as described in the following. In the case of the spleen ultrasound image, since the size of the spleen is much smaller than the size of the liver and since the spleen generally doesn't exhibit average brightness variation, the parameters are extracted in only one ROI. Two examples of such sets of ROIs with the corresponding histograms at the three depths in two ultrasound hepatic images – one corresponding to a healthy liver and the other exhibiting steatosis – are given in Figure 2. Examining the histograms, one observes that the shape and position on the grey scale axis of the three histograms (at the three depths) is the same in the case of the healthy liver (Figure 2.a)). However in the case of the ultrasound image corresponding to severe steatosis, the shape and position of the grey level histograms modifies with the increase depth: not only the median grey level value decreases, but also the spread and the skew of the histograms change from a depth to another (Figure 2.b)).

2)

3)

4)

5)

6)

a)

its values at the three different depths in the hepatic ultrasound image. With these three values we form the 3-D feature space of the 10th percentile, used in the classification phase, described by the vectors l10 in the form: l10=[l10/0.5 l10/2 l10/3.5]T. The 90th percent percentile of the normalised grey level histogram, which is a descriptor of the brightest grey level in the ROI. l90/0.5, l90/2 and l90/3.5 denote its values at the three different depths in the hepatic ultrasound image, and its corresponding feature space, used in the classification phase, is described by the vectors l90 in the form: l90=[l90/0.5 l90/2 l90/3.5]T. The median of the normalised grey level histogram, which is a descriptor of the average grey level in the ROI. ml/0.5, ml/2 and ml/3.5 denote its values at the three different depths in the hepatic ultrasound image. With these values, the 3-D feature space of the median grey level, used in the classification phase, is described by the vectors ml in the form: ml=[ ml/0.5 ml/2 ml/3.5]T. The median of the normalised grey level histogram referenced to the median of the spleen, computed by subtracting from the hepatic grey level median vector ml, the median of the normalised grey level histogram in the ROI section of the spleen, ms. The corresponding feature space is formed by the vectors ml-s=[ml/0.5-ms ml/2-ms ml/3.5-ms]T. The standard deviation of the normalised grey level histogram, which can describe the grey levels spread. sdl/0.5, sdl/2 and sdl/3.5 are its values at the three different depths. With these three values we form the feature vectors sdl: sdl=[sdl/0.5 sdl/2 sdl/3.5]T. The skew of the normalised grey level histogram, which can describe the symmetry of the grey levels distribution in respect to their median. Denoting the ROI width W and the ROI height H (in pixels), we compute the skew of the normalised grey level ROI histogram with the formula:

b) Figure 1. The software application for ultrasound image analysis interface: a) ROI positioning; b) parameter extraction window.

The features used here, as locally global descriptors of intensity distribution, are the following: 1) The 10th percent percentile of the normalised grey level histogram, which is a descriptor of the darkest grey level in the ROI. l10/0.5 (the 10th grey level percentile at the depth 0.5 cm), l10/2 and l10/3.5 denote

a)

b)

Figure 2. Example ROIs and histograms with the intensity in-depth variation in ultrasound hepatic images: (a) normal tissue; (b) severe steatosis.

219

1st International Conference on Advancements of Medicine and Health Care through Technology, MediTech2007, 27-29th September, 2007, Cluj-Napoca, ROMANIA perform the SVM classifier training. To do so, one must define a training set of patterns, i.e. in our case, a set of labelled tissue samples containing both examples of healthy tissue and examples of tissue with severe steatosis, as balanced as possible in respect to the number of examples in both classes. The tissue samples and labels are common to all the feature spaces, only the features differ from one space to another. Several SVM classifiers were investigated, with different parameters and kernel functions: the linear classifier; the non-linear SVM with polynomial kernel function of degree 2, 3, 5; the non-linear SVM with Gaussian RBF kernel function, for different values of the width of the Gaussian. The penalty factor C assigned to the classification errors was also varied: 1, 10, 100, 1000. Finally, the classifier giving the best separation in the training set of our study group (as described in the Experimental Results section) in most of the feature spaces was selected, that is, the non-linear SVM with Gaussian RBF kernel, with the width of the Gaussian kernel of 0.001 and the penalty factor C=100. The optimality of the Gaussian RBF kernel based SVM classifiers among other choices is often encountered in medical pattern classification applications [14, 15].

3 ⎛ (l − m l ) ⎞ ⎞⎟ 1 ⎛⎜ 255 ⎟ (1) h(l) ⋅ ⎜⎜ ⎟ ⎟ W ⋅ H ⎜ l =0 ⎝ sd l ⎠ ⎠ ⎝ where h:{0,1,...,255}→[0;1] is the normalised grey level histogram in the currently considered ROI. We denote the skew values at the three considered depths by sl/0.5, sl/2 and sl/3.5 and accordingly, define the skew feature vectors as sl=[sdl/0.5 sdl/2 sdl/3.5]T. 7) The coefficient of variation of luminance (CVL) of the normalised grey level histogram is the last examined quantitative descriptor, given by: sd (2) cvl = l ⋅100[%] l avg

sl =



where lavg is the average (mean) of all the grey levels in the currently considered ROI. We denote the values of CVL at the three considered depths by cvl0.5, cvl2 and cvl3.5 and accordingly, define the feature vectors cvl=[cvl0.5 cvl2 cvl3.5]T. Considering these seven features extracted from ROIs placed at the three above-mentioned depths for each hepatic US image, we will have 7 feature spaces to examine with the pattern classifier, describing the local intensity distribution and its in-depth variation.

A rather common situation in binary pattern classification in low dimensional feature spaces is the one when although the correct classification rate (accuracy) is not very high in each individual space, the errors are unevenly distributed in the different spaces: samples incorrectly classified in some feature spaces are correctly classified in the others. In such a case, the overall classification accuracy can be improved either by fusing the individual feature spaces into a new joint feature space, or by performing a classification independently on each individual feature and fusing the individual classifiers decisions by some decision combination scheme [16]. Depending on the specific application and on the particularity of the extracted individual features, one can choose the more suitable of the two strategies. In the case of the features considered here for the discrimination between the tissue with no steatosis and with severe steatosis, the fusion of classifiers’ decisions proved more appropriate than the fusion of the feature spaces. This can be accounted for the different nature of some of the extracted features (e.g. the median and the percentiles as compared to the skew and CVL of the normalised grey level histogram of the ROI). We use one of the most simple classifier decision fusion strategies, namely, the majority vote [16], in which the final classification decision is adopted as follows. Let us consider a certain ultrasound hepatic image (to be classified as showing severe steatosis or no steatosis). For this image’s classification, we will denote the seven corresponding individual SVM classifiers binary decisions (in the seven feature spaces) as fp10(l10), fp90(l90), fmed(ml), fmed-rell(ml-s), fsd(sdl), fskew(sl) and fcvl(cvl), with values in the set {-1,+1}, with the decision -1 if the currently examined ultrasound image exhibits severe steatosis and +1 if it represents a healthy liver. Then the overall classification decision by the majority voting

2.2. Tissue Classification with Support Vector Machines Support vector machines (SVMs) are powerful binary supervised classifiers, from the class of machine learning techniques, based on the optimal separating hyper-plane principle between the two data classes to be identified [13]. Due to their mathematically proven and applications validated ability of learning in sparse feature spaces from relatively few training examples, with good generalisation ability, SVM classifiers are successfully used nowadays in various pattern recognition applications, including medical applications as the one described in this paper (e.g. [9]). One of the reasons of preferring SVMs to solve binary classification problems is their ability to derive an optimal separation hyper-plane even for data that are not linearly separable in their original feature space, by projecting these data in higher dimensional feature space and deriving a linear separation surface in this higher dimensional space (non-linear SVMs). Furthermore, in deriving and expressing the equation of the decision surface (optimal separating hyper-plane), the higher-dimensional feature space does not need to be explicitly used. The decision equation can be expressed as a function of the original vectors’ dot product using some suitably chosen kernel functions, thus yielding the computations very simple [13]. The hepatic tissue classification considered in this work can be expressed as a binary classification problem with the following two categories as classes: healthy hepatic tissue and hepatic tissue with severe steatosis. As feature spaces used to derive and employ the binary SVM tissue classifier, we investigate each of the seven spaces described in the previous section individually. As in any SVM pattern classification problem, the first step is to 220

1st International Conference on Advancements of Medicine and Health Care through Technology, MediTech2007, 27-29th September, 2007, Cluj-Napoca, ROMANIA scheme is given as:

number of positive test patterns by N + t and the number

7 ⎧ ⎪+ 1, if f k (x k ) > 0 (3) f maj (x ) = ⎨ k =1 ⎪⎩− 1, otherwise where fk denotes one of the seven individual classifiers binary decision functions and xk denotes one of the seven individual extracted feature vectors for the currently analysed ultrasound image (e.g. f1=fp10, x1=l10), and x refers generically the current ultrasound hepatic image.



of negative test patterns by N − t . Also we denote the number of positive test patterns falsely classified as negative by FN and the number of negative test patterns falsely classified as positive by FP . With these notations, the accuracy is: Accuracy [%] = (FP + FN ) N t ⋅100 (4) The accuracies given by each feature space with the corresponding SVM classifier are given in Table 1.

3. EXPERIMENTAL PROTOCOL

Table 1. The discrimination performance of the seven feature spaces, for steatosis classification, with individual SVM classifiers

The verification of the proposed ultrasound image analysis method described in the previous section was done on a study group of 62 patients (30 women and 32 men) with chronic viral hepatitis B, C and non-alcoholic steatohepatitis, enrolled in our study in the period October 2005 - July 2006. The diagnostic was established by standard methods, including histopathological liver examination. Each patient went through hepatic and spleen ultrasonography in the same image acquisition conditions, performed on the same ultrasound machine by the same examiner. We used the spleen section as reference ultrasound tissue sample for the patient himself. Later on, the patients went through liver biopsy. The fragment was histopathologically examined with hematoxilin-eosine staining. The severity degree of steatosis was quantified using the Brunt classification [7]. Histopathologically, the subjects were classified in: 40 without steatosis or mild steatosis and 22 with steatosis degree 2-3 (severe steatosis).

Feature l10 l90 ml ml-s sdl sl cvl

Accuracy[%] 75 71.88 71.88 85.19 65.63 62.5 65.63

Discussing the results in Table 1, we notice that, as expected, the best classification performance is given by the relative median hepatic luminance feature space, referenced to the spleen median luminance, ml-s. Therefore for this feature vector, we consider useful to compute and present also other measures of performance, currently used in the evaluation of similar ultrasound image analysis applications, namely: 1) the sensitivity of the method, defined as: Sensitivity [%] = 1 − FN N + t ⋅100 . (5) 2) the specificity of the method, defined as: Specificity [%] = 1 − FP N − t ⋅100 . (6) 3) the predictive positive value (PPV): PPV [%] = N + t − FN N + t − FN + FP ⋅100 (7) 4) the predictive negative value (PNV): PNV [%] = N − t − FP N − t − FP + FN ⋅100 (8) For the median of the normalised grey level histogram referenced to the median of the spleen, these performances are presented in Table 2.

(

4. EXPERIMENTAL RESULTS In evaluating the performance of the method proposed here, we use the study group of 62 individuals mentioned above. From the total of 40 subjects with mild or no steatosis and 22 subjects with severe steatosis, a sub-set of ultrasound images (a pair of images spleen – liver per subject) perfectly balanced, corresponding to 30 patients, 15 with severe steatosis and 15 with no mild or no steatosis, are used as training set of ultrasound images. In each of the seven feature spaces considered for examination, the extracted feature vectors l10, l90, ml, ml-s, sdl, sl and cvl from the 15 ultrasound images with severe steatosis are labelled as -1, whereas the same feature vectors extracted from the 15 ultrasound images with mild or no steatosis are labelled as +1. The remaining sub-set of 32 ultrasound images, 25 with mild or no steatosis and 7 with severe steatosis, are used as test set.

(

)

)

((

)(

))

((

)(

))

Table 2. Discrimination performances between severe and mild/no steatosis in the relative median luminance feature space (using as reference the spleen)

Performance Sensitivity Specificity Accuracy PPV PNV

The first set of experiments aims to assess the ability of each of the seven feature spaces to discriminate correctly and reliably between the cases with severe steatosis and the ones with mild or no steatosis, using the individually trained nonlinear SVM classifiers with Gaussian RBF kernel. Since the classification error in the training set was 0, we present here as the results of this experiment only the classification accuracy in the test set. Here we denote the total number of test patterns by N t , the

In training set [%] 100 100 100 100 100

In test set [%] 95.65 25 85.19 88 50

One can notice very good rates for the sensitivity, accuracy and PPV; however the values of the specificity and the PNV are not satisfactory. This can be partially accounted for the few number of negative test examples 221

1st International Conference on Advancements of Medicine and Health Care through Technology, MediTech2007, 27-29th September, 2007, Cluj-Napoca, ROMANIA available (only 7), which reduces the statistical significance of the test in this case.

CNCSIS research project 1414/2007, financed by the Romanian Government.

The second experiment performed started from the observation (validated by the first set of experiments) of an uneven distribution of errors in the test set among the seven SVM classifiers – which makes it reasonable to expect improved classification performance of the ultrasound hepatic images into severe steatosis and mild or no steatosis through the fusion of the individual decisions of these binary classifiers, as e.g. using the majority vote as described in the previous section. Indeed, applying this classifier decision fusion strategy, the overall performance of the system is improved, as shown by the results in Table 3. Especially remarkable is the improvement of the specificity (from 25% in the case of using only the relative median luminance feature space to 85.71% when using the majority vote classifier fusion scheme); this makes acceptable the slight decrease in the sensitivity (from 95.65% to 88%), especially considering that all the other performances are better in the case of classifier decisions combination.

7. REFERENCES [1] Schiff E.R, Sorrell M.F., Schiffs disease of the liver, 10th Ed. Lippincot, Williams & Wilkins 2007. [2] Martin L.M., Sheridan M.J., Younossi Z.M., The impact of liver disease on health-related quality of life: a review of the literature, Curr Gastroenterol Rep, pp. 79-83, 2002. [3] Brunt E.M., Nonalcoholic steatohepatitis: definition and pathology, Semin Liver Dis., pp.3-16, 2001. [4] Badea R.I., Ficatul. In: R.I. Badea, S.M. Dudea, P.A. Mircea & F. Stamatian (eds). Tratat de ultrasonografie clinica, vol.1. Ed. Medicala, Bucuresti 2000. [5] Mathiese U.L., Franzen L.E., Aselnis H., Resjo M., Increased liver echogenicity at ultrasound examination reflects degree of steatosis but not of fibrosis in asymptomatic patients with abnormalities of liver transaminases, Dig Liver Dis, pp. 516-22, 2002. [6] Saadeh S., Younossi Z.M., & Remer E.M., The utility of radiological imaging in non-alcoholic fatty liver disease, Gastroenterology, pp. 745-50, 2002. [7] Brunt E.M., Janney C.G., Di Bisceglie A.M., Neuschwander-Tetri B.A., Bacon B.R., Non-alcoholic steatohepatitis: a proposal for grading and staging the histological lesions, Am J Gastroenterol, pp. 2477-74, 1999. [8] Kadah Y.M., Farag A.A., Zurada J.M., Badawi A., Youssef A. Classification Algorithms for Quantitative Tissue Characterization of Diffuse Liver Disease from Ultrasound Images, IEEE Transactions on Medical Imaging, Vol. 15, No. 4, pp. 466-478, August 1996. [9] Cao G.T., Shi P.F., Hu B., Liver fibrosis identification based on ultrasound images captured under varied imaging protocols, J Zhejiang Univ Sci B., Nov 6(11), pp. 1107-1114, 2005. [10] Mojsilovic A., Markovic S., Popovic M., Characterization of visually similar diffuse diseases from B-scan liver images with the nonseparable wavelet transform, ICIP, pp. 547-550, 1997. [11] Lee C., Choi J., Kim K., Seo T., Lee J., Park C., Usefulness of standard deviation on the histogram of ultrasound as a quantitative value for hepatic parenchymal echo texture; preliminary study, Ultrasound in Medicine & Biology, Vol. 32, No. 12, pp. 1817-1826. [12] Pratt W. K., Digital Image Processing: PIKS Inside, 3rd Edition, John Wiley & Sons, 2001. [13] Vapnik V.N., Statistical Learning Theory, J. Wiley, N.Y., 1998.

Table 3. Discrimination performances between severe and mild/no steatosis using the majority voting scheme to fuse the individual binary decisions of the SVM classifiers

Performance Sensitivity Specificity Accuracy PPV PNV

In training set [%] 100 100 100 100 100

In test set [%] 88 85.71 87.5 95.65 66.66

As compared to the state of the art in the literature, the results presented in Table 3 outperform some of the performances obtained in feature spaces with similar computational complexity; thus, the reported accuracies in discriminating severe steatosis from mild or no steatosis are [10]: using as features some fractal measures– 69%; using features computed from the Fourier power spectrum – 82% (both inferior to our accuracy – of 87.5%). The grey level co-occurrence matrix-based features slightly outperform our accuracy (87%) and so do the non-separable quincunx wavelet transform decomposition features (90% accuracy), however at a higher computational cost.

[14] Tsantis S., Cavouras D., Kalatzis I., Piliouras N., Dimitropoulos N., Nikiforidis G., Development of a support vector machine-based image analysis system for assessing the thyroid nodule malignancy risk on ultrasound, Ultrasound in Medicine and Biology Vol. 31(11), pp. 1451-1459, 2005. [15] Kyriacou E., Pattichis C.S., Pattichis M.S., Mavrommatis A., Panagiotou S., Christodoulou C.I., Kakkos S., Nicolaides A., Classification of Atherosclerotic Carotid Plaques Using Gray Level Morphological Analysison Ultrasound images, Proc. 3rd IFIP Conference on Artificial Intelligence Applications & Innovations (AIAI)06, Athens, pp.737-744, Greece, 2006.

5. CONCLUSION We investigated and proposed a new method for the quantitative evaluation of hepatic steatosis from ultrasound images. The experimental results are close to the state of the art in the field, while using very simple tissue descriptors. The relative median hepatic luminance using as reference the spleen has a rather high accuracy in steatosis quantification, and this accuracy can be increased with the support of the other parameters.

[16] Ruta D., Gabrys B., An Overview of Classifier Fusion Methods, Computing and Information Systems, Univ. Paisley, ISSN 1352-9404, Vol. 7, No. 1, pp. 1-10, February 2000.

6. ACKNOWLEDGMENTS This work was performed in the framework of the 222