An integrated index for identification of fatty liver disease using radon transform and discrete cosine transform features in ultrasound images

Accepted Manuscript

An Integrated Index for Identification of Fatty Liver Disease Using Radon Transform and Discrete Cosine Transform Features in Ultrasound Images U Rajendra Acharya , Hamido Fujita , Vidya K Sudarshan , Muthu Rama Krishnan Mookiah , Joel EW Koh , Jen Hong Tan , Yuki Hagiwara , Chua KC , Junnarkar Sameer Padmakumar , Anushya Vijayananthan , Kwan Hoong Ng PII: DOI: Reference:

S1566-2535(15)00119-0 10.1016/j.inffus.2015.12.007 INFFUS 760

To appear in:

Information Fusion

Received date: Revised date: Accepted date:

27 April 2015 18 November 2015 22 December 2015

Please cite this article as: U Rajendra Acharya , Hamido Fujita , Vidya K Sudarshan , Muthu Rama Krishnan Mookiah , Joel EW Koh , Jen Hong Tan , Yuki Hagiwara , Chua KC , Junnarkar Sameer Padmakumar , Anushya Vijayananthan , Kwan Hoong Ng , An Integrated Index for Identification of Fatty Liver Disease Using Radon Transform and Discrete Cosine Transform Features in Ultrasound Images, Information Fusion (2015), doi: 10.1016/j.inffus.2015.12.007

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Highlights Classification of Fatty Liver Disease and normal images is proposed. Radon Transform and Discrete Cosine Transform method are used. Index from features is developed to discriminate classes using a single number. Proposed system obtained an average accuracy of 100%.

AC

CE

PT

ED

M

AN US

CR IP T

• • • •

1

ACCEPTED MANUSCRIPT

An Integrated Index for Identification of Fatty Liver Disease Using Radon Transform and Discrete Cosine Transform Features in Ultrasound Images U Rajendra Acharya1,2,3*, Hamido Fujita4, Vidya K Sudarshan1, Muthu Rama Krishnan Mookiah1, Joel EW Koh1, Jen Hong Tan1, Yuki Hagiwara1, Chua KC1,

CR IP T

Junnarkar Sameer Padmakumar5, Anushya Vijayananthan6, Kwan Hoong Ng6 Department of Electronics and Computer Engineering, Ngee Ann Polytechnic,

1

Singapore, 599489

Department of Biomedical Engineering, School of Science and Technology, SIM

2

AN US

University, Singapore, 599491

Department of Biomedical Engineering, Faculty of Engineering, University of

3

Malaya, Malaysia, 50603

Iwate Prefectural University (IPU), Faculty of Software and Information Science,

4

M

Iwate, Japan, 020-0693

Hepatobiliary and Pancreatic Surgery, Tan Tock Seng Hospital, Singapore, 308433

ED

5

Department of Biomedical Imaging, Faculty of Medicine, University of Malaya,

PT

6

Malaysia, 50603

CE

*Corresponding Author

AC

Postal Address: Department of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore 599489 Telephone: (65) 64606135; Email Address: [email protected] (Acharya UR)

Abstract

Alcoholic and non-alcoholic fatty liver disease is one of the leading causes of chronic liver diseases and mortality in Western countries and Asia. Ultrasound image assessment is most commonly and widely used to identify the Non-Alcoholic Fatty Liver Disease (NAFLD). It is one of the faster and safer non-invasive methods of 2

ACCEPTED MANUSCRIPT

NAFLD diagnosis available in imaging modalities. The diagnosis of NAFLD using biopsies is expensive, invasive, and causes anxiety to the patients. The advent of advanced image processing and data mining techniques have helped to develop faster, efficient, objective, and accurate decision support system for fatty liver disease using ultrasound images. This paper proposes a novel feature extraction models based on Radon Transform (RT) and Discrete Cosine Transform (DCT). First

CR IP T

Radon Transform (RT) is performed on the ultrasound images for every 1 degree to capture the low frequency details. Then 2D-DCT is applied on the Radon transformed image to obtain the frequency features (DCT coefficients). Further the 2D-DCT frequency coefficients (features) obtained are converted to 1D coefficients

AN US

vector in zigzag fashion. This 1D array of DCT coefficients are subjected to Locality Sensitive Discriminant Analysis (LSDA) to reduce the number of features. Then these features are ranked using minimum Redundancy and Maximum Relevance (mRMR) ranking method. Finally, highly ranked minimum numbers of features are

M

fused using Decision Tree (DT), k-Nearest Neighbour (k-NN), Probabilistic Neural Network (PNN), Support Vector Machine (SVM), Fuzzy Sugeno (FS) and AdaBoost

ED

classifiers to get the highest classification performance. In this work, we have obtained an average accuracy, sensitivity and specificity of 100% in the detection of

PT

NAFLD using FS classifier. Also, we have devised an integrated index named as

CE

Fatty Liver Disease Index (FLDI) by fusing two significant LSDA components to distinguish normal and FLD class with single number.

AC

Keywords: Ultrasound, Radon transform, Discrete cosine transform, Locally sensitive discriminant analysis, Fatty liver disease, Fuzzy classifier, AdaBoost.

1. Introduction Fatty Liver Disease (FLD) or hepatic steatosisis the excessive accumulation of lipids (triglycerides) in the hepatocytes (liver). It is named as non-alcoholic FLD (NAFLD) in the absence of excessive alcohol consumption [1] and is the leading chronic liver

3

ACCEPTED MANUSCRIPT

disease worldwide. Diabetes and obesity are the major risk factors of NAFLD [28]. The estimated prevalence of NAFLD in adult population is around 15%-30%, which increases with age [2-4]. It is estimated that the prevalence of NAFLD is 32% of general population in India [5] and 15% in China [6]. Prevalence of NAFLD in diabetics is 74% in North America and 70% in Italy [7, 8] due to obesity in Asia

CR IP T

ranging from 50% to 80% in Japan, 70-80% in China, 10-50% in Korea [9-13]. NAFLD pathology may be accompanied with swelling of liver or may advance to an extreme situation, liver cirrhosis, leading to permanent liver damage [14, 15]. Nevertheless, the condition may be reversible if diagnosed in its early stage [15, 16].

AN US

It may lead to fibrosis [17], cirrhosis [18], liver cancer [19, 20], liver failure requiring liver transplant [21], and mortality [22]. The pathological changes introduced by this disease can be evaluated using B-mode ultrasound (US) images [23, 24]. However, it is associated with the shortcomings such as inter-operator variability, subjective

M

evaluation, and restricted potential to measure the amount of fatty infiltration. Few qualitative reviews [25, 26] have challenged the ability of ultrasound to accurately

ED

detect fatty liver.

Various computer-aided techniques are proposed for automatic detection of FLD

PT

using ultrasound images [27]. Yasser et al., (1996) [28] used gray level features, attenuation and back scattering features for the detection of FLD and reported a

CE

sensitivity of 100% using 9 features. Pavlopoulos et al., (2000) [29] used texture

AC

features namely Fractal Dimension (FD), Spatial Gray-Level Dependence Matrices (SGLDM), Gray-Level Run Length Statistics (GLRLS), Gray-Level Difference Statistics (GLDS) and First-Order gray-level Parameters (FOP) to discriminate normal and FLD classes. Their method obtained an accuracy of 82.67% using 12 features. Yen et al. (2004) [16] proposed automated diagnosis of FLD using GrayLevel Co-occurrence Matrix (GLCM) and Non-Separable Wavelet Transform (NSWT). Their method used 10 features and obtained an accuracy of 90.5%.

4

ACCEPTED MANUSCRIPT

Lupsor et al., (2011) [30] compared two computer-aided techniques namely, the attenuation coefficient and first order texture features for the FLD (steatosis) assessment using US images. Andreia et al., (2012) [31] used Gray-Level Cooccurrence Matrix (GLCM), gray level run length matrix (GLRLM), FD and Law Texture Energy (LTE) to extract the total of 325 features from ultrasound images. Three classifiers (AN, SVM and KNN) are used for the classification of liver steatosis

CR IP T

(FLD). Their study showed that the SVM classifier achieved an accuracy of 79.77% compared to other classifiers tested.

Ricardo et al., (2009, 2012) [32, 33] proposed CAD system for steatosis analysis

AN US

and classification using texture feature extraction method. The texture features (attenuation coefficient and Auto Regressive (AR)) are extracted using Discrete Wavelet Transform (DWT) detail coefficients of US images. The Bayes classifier achieved an overall accuracy of 93.54% and 95% for the classification of normal and

M

steatosis classification using AR coefficients features extracted from the multi-scale Haar wavelet decomposed US images. Singh et al., (2012, 2014) [34, 35] developed a

ED

new quantitative method for liver classification (normal and FLD) using ultrasound images. Their study used five different texture feature extraction methods (i) spatial

PT

gray-level dependence matrices, (ii) statistical feature matrix, (iii) Law’s Texture Energy Measures (TEM), (iv) Fourier power spectrum (FPS) and (v) fractal features

CE

from the ultrasound images. The study used statistical and linear discrimination analysis to select the best features and achieved an accuracy of 92% and 95% using 6

AC

and 7 texture features respectively. Acharya et al., 2012a [36] presented a CAD technique (called Symtosis) for the

automated detection of FLD using texture, DWT, and HOS bispectrum features extracted from ultrasound images. They have reported an average accuracy of 93.3% using decision tree classifier. Minhas et al., (2012) [37] presented a novel approach for detection of FLD using texture analysis of liver ultrasound images. Their study used multiscale analysis capability of Wavelet Packet Transform (WPT) to extract 5

ACCEPTED MANUSCRIPT

statistical features which changes in echogenicity, granularity and homogeneity of ultrasound due to the incidence of FLD. Their method claimed classification accuracy of 95% using leaner SVM classifier. Recently, Dan et al., (2013) [38] discussed and compared the CAD method for steatosis rating (severity) in ultrasound images using Radon Forest (RF) and SVM classifiers.

CR IP T

RT-DCT based approach is used for face recognition [39]. However, the available literatures show that there is no study used RT-DCT based approach for automated fatty liver disease detection. Moreover, RT-DCT combination captures the subtle changes in the images [39]. Hence, this paper proposes a novel approach for an

AN US

automated detection and classification of ultrasound FLD using Radon Transform (RT), Discrete Cosine Transform (DCT), and Locality Sensitive Discriminant Analysis (LSDA) method [39]. In this method, the ultrasound images of normal and FLD are pre-processed to enhance the image contrast and subjected RT for every 1°

M

angle to obtain the Radon projections from 0-179°. DCT is performed on those Radon projections to obtain the DCT coefficients matrix. DCT coefficients matrix which is

ED

the frequency coefficients are arranged in zigzag manner to obtain a vector features arranged in the increasing frequency arrangement. Hence, these DCT coefficients are

PT

converted to 1D array of coefficients. All the 1D DCT features are subjected to LSDA data reduction method to reduce the number of features to 30. These reduced

CE

numbers of features are ranked using minimum Redundancy and Maximum Relevance (mRMR) ranking method. Highly ranked features are fed to Decision Tree

AC

(DT), k-Nearest Neighbour (k-NN), Probabilistic Neural Network (PNN), Support Vector Machine (SVM), Fuzzy Sugeno (FS) and AdaBoost classifiers one by one to get the highest classification performance using optimum features. Detailed descriptions of each step are explained in the following sections.

2. Data Used One hundred ultrasound liver images were used to evaluate the performance of our

6

ACCEPTED MANUSCRIPT

methodology. Figure 1 shows the normal and FLD ultrasound images. Among these 100 cases, 50 were abnormal (affected by FLD) and 50 were normal images. The ultrasound images of normal and fatty livers were acquired by expert operators with the ultrasound equipment in University of Malaya Hospital, Malaysia. All the images were collected from routine cases and were consecutively recruited. The ethical clearance was obtained from the ethical review committee of University of

CR IP T

Malaya Hospital, Malaysia. The ultrasound images were obtained by a Philips CX© 50 ultrasound machine. All images were captured with 1898×888 and 1418×720 pixels with a grey level resolution of 8 bits/pixel. Images were stored in the Digital Imaging and Communications in Medicine (DICOM) format. The broadband curved

AN US

array transducer C5-1 from Philips© was used. It is composed by 160 piezoelectric elements with a curved array shape, and had the operating frequency range from 1

AC

CE

PT

ED

M

to 5 MHz.

Figure 1: Normal liver image (left column) and abnormal liver image (right column).

The ultrasound machine is preset for this study, after the calibration step. Ultrasound is calibrated at 3.5 MHz frequency, an image depth of 15cm. The dynamic range was set to 70 dB with variable gain. Time gain compensation (TGC) maintained constant by setting it to its central position throughout the procedures, eliminating this variable parameter. The ground truth of each image whether normal

7

ACCEPTED MANUSCRIPT

or abnormal was determined manually by the operators and confirmed by laboratory analysis.

3. Methodology This paper used a new method of automated detection of NAFLD in which DCT frequency features are extracted from the RT of the images. The block diagram of an

CR IP T

automated FLD detection system is shown in Figure 2. The current method comprises of four stages: (i) in the first stage, the ultrasound images are preprocessed, (ii) in the second stage, the RT is performed for every 1° angle to obtain the RT projections of an image, (iii) in the third stage, the DCT is used on the whole

AN US

RT projections to acquire the DCT coefficients i.e. to extract the rotation invariant frequency features, and (iv) in the last stage, few of the DCT coefficients are selected to assemble feature vectors i.e. conversion of 2D DCT coefficients into 1D array of

CE

PT

ED

the methods used is as follows.

M

DCT features, using deterministic approach (zigzag approach). A brief description of

system.

AC

Figure 2: Block diagram of an automated FLD detection

3.1 Pre-processing From each image, the patient information is removed and images have various pixel resolutions such as 1898×888 and 1418×720. Hence, it is standardised to 500×500 pixels resolution. Further, resized US images are enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE) [40]. It divides the images into small blocks and perform adaptive histogram equalization on these blocks [40]. Hence, 8

ACCEPTED MANUSCRIPT

pixel intensities of these blocks mapped into new intensity values which is proportional to rank of each block intensity histograms [40]. It enhances the contrast of normal and FLD ultrasound images. 3.2 Radon Transform (RT)

(

)

Where

[ (

)]

∫

∫

(

)is defined as,

)(

CR IP T

The radon transform (RT) of a 2-D function (

)

is the distance of a line from the origin and

[

(1)

] is the angle between

the distance vector and x-axis [41, 42]. The symbol

denotes RT operator. A rotation

of (

) in the variable

leads to translation of (

by . Benefit

AN US

) by an angle

of using RT, in the current approach, is robustness to zero mean white noise. Figure

CE

PT

ED

M

3 shows the RT of original image for 0-179° orientations.

(b)

AC

(a)

Figure 3: Radon transform of image: (a) normal (Figure 2(a)), (b) NAFLD (Figure 2(b)).

3.3 Two Dimensional Discrete Cosine Transform (2D-DCT) The DCT is a prevailing method used for image compression, which transforms images into frequency representation from the spatial form. Like other transforms, DCT also provides energy compaction. The DCT coefficients [43] of an (

) is given by, 9

image

ACCEPTED MANUSCRIPT ( ) ( )∑

∑

(

)

(

(

)

(

) is a 2D matrix of DCT coefficients.

)

(

)

)

(

)

(2)

)

∑

∑

( ) ( ) (

)

Where, ( )

( )

√

) is the average value of an image, also known as the DC coefficient. The

ED

(

√

{ √

(3)

M

√ {

(

AN US

(

CR IP T

And the inverse DCT is given by,

cosines of DCTs are orthogonal to each other. In this work, DCT is applied on RT

PT

projections of whole image to derive the frequency feature (coefficients) matrix [43] which reflects the FLD changes in the US images.

CE

3.4 DCT coefficients selection

AC

Once the DCT coefficients matrix is obtained from the RT projections, only few coefficients are selected while leaving the others. The DCT coefficients selection is a significant stage of the feature extraction process. Generally, conventional methods – zigzag or zonal masking – are used for the coefficients selection. In this paper, significant coefficients of DCT are concatenated to form the image feature vector using zigzag approach i.e. conversion of 2D DCT coefficients into 1D array of DCT features. Figure 4 (a) displays the DCT coefficients matrix with zigzag 10

ACCEPTED MANUSCRIPT

approach used in this work to select the DCT coefficients. In the DCT coefficients matrix, the top left corner coefficients, carry most of the image energy and depicts the mean of the whole matrix. The remaining coefficients depict the variations of intensity among the block images. Figure 4 (b) shows the array of 1D DCT coefficients selected or extracted using zigzag approach from the 2D DCT

ED

M

AN US

approach to a 1D array of DCT coefficients (Figure 4).

CR IP T

coefficients matrix. In other words, 2D DCT coefficients are converted using zigzag

PT

Figure 4: (a) DCT coefficients selection using zigzag approach, and (b) 1D array of DCT coefficients obtained from (a).

CE

3.5 Feature Reduction and Ranking

AC

In this work, 1D DCT coefficients are obtained from the radon transformed images. We have extracted a total of 127,980 1D-DCT coefficients from 2D DCT matrix in zigzag fashion. So, these features are applied to LSDA method for dimensionality reduction yielding 30 LSDA components. 3.5.1 Locality Sensitive Discriminant Analysis (LSDA) Brief description of LSDA algorithm is given [44] below.

11

ACCEPTED MANUSCRIPT Consider n data samples x1 , x2 ,........, xn  R a sampled from the underlying sub manifold M. Then the two nearest neighbour graphs in LSDA: the within-class graph

G w and the between-class graph Gb can be constructed. For each data points x i , its k nearest neighbours are determined; then an edge between x i and its neighbours is drawn.

1, WTw   0,

if xi  N w ( x j ) or x j  N w ( xi )

1, WTb   0,

if xi  N b ( x j ) or x j  N b ( xi ) otherwise

(4)

(5)

AN US

otherwise

CR IP T

Weight matrices WTw and WTb of G w and Gb are defined as,

Where, the set N w ( xi ) contains the points which are the k nearest neighbours of x i and share the same label xi . Whereas N b ( xi ) contains the points which are the k

M

nearest neighbours of x i and share the different label x i . It is clear to see

G w and Gb : G  Gw  Gb

ED

WT  WTw  WTb and the nearest neighbour graph G can be identified as a mixture of

PT

Now consider the G w and Gb graph to line mapping complication so that connected points of G w stay as near together as possible while connected points of Gb stay as far

CE

as possible.

AC

In LSDA, the within-class compactness can be learned by Laplacian embedding, i.e.,

Gw  min  ( yi  y j )2WTw

(6)

i, j

The between-class separability can be learned by

Gb  max  ( yi  y j )2WTb

(7)

i, j

12

ACCEPTED MANUSCRIPT Where yi   T xi and y j   T x j is the one-dimensional representation of xi and x j . A logical criterion for selecting a good map is to optimize the two objective functions – equations (6) and (7). 3.5.2 minimum Redundancy and Maximum Relevance (mRMR)

CR IP T

LSDA components are ranked to select unique and highly discriminating features using mRMR ranking technique. This method uses maximum relevance and minimum redundancy measures to rank the features [45, 46]. Maximum relevance search features having high mutual information with target class. However, some features may show high discrimination on the target class due to high correlation.

AN US

Hence, minimum redundancy criteria is used to find features without correlation [45, 46]. These two criteria are combined to get optimum features. In this work, mutual information is computed using fuzzy entropy [47] and mRMR-mutual information difference method to rank LSDA components. Further, ranked LSDA

M

components are fed to various supervised classifiers to obtain classification

3.6 Classification

ED

performance using 10-fold cross validation method.

PT

Classification is an important process in the development of automated detection system. In this process, highly ranked significant features are subjected to classifiers

CE

for the automated classification into normal and FLD classes. For this classification

AC

purpose, classifiers namely DT, k-NN, PNN, SVM, FS and AdaBoost are used. The brief descriptions of each classifier are explained below. Decision Tree (DT): This classifier using the selected features constructs a tree from the training data [48]. The rules are obtained from this constructed tree which is used to classify the two classes, and these rules are used to determine the class of the test data. Better the design of the constructed tree better is the performance of this

13

ACCEPTED MANUSCRIPT

classifier. In this work, Classification And Regression Tree (CART) algorithm [48] is implemented to perform classification. k-Nearest Neighbour (k-NN): It is a simple classifier that determines the k-nearest neighbors by using the minimum distance from the testing and training data [49, 50]. The most common among the k-nearest neighbors are assigned with a class. This

CR IP T

classifier has poor run-time performance when training set is large. In this work, k = 5 is used [49, 50].

Probabilistic Neural Network (PNN): It is a multi-layered feedforward neural network which can be used for classification. It uses supervised learning algorithm

AN US

to compute the weights. It has four layers: input, pattern, summation and output [51]. It require short training time and sensitive to outliers. The limitation of this classifier is that it takes long time to test the class of the unknown data and requires

M

lot of memory while training.

Support Vector Machine (SVM): SVM constructs an (N-1)-dimensional hyperplane

ED

that classifies the data into two categories where N represents the number of input features. The hyperplane acts as a decision surface separating two classes with a

PT

maximum margin [52, 53]. The vectors near the hyperplane on either side are called support vectors. The training set consists of a target variable called class label and

CE

remaining variables are called feature variables. The generated SVM model maps the features of a test data on the same space as that of training data features and predicts

AC

the class of a test data [54, 55]. Fuzzy Sugeno (FS): It consists of Fuzzy Inference System (FIS) with fuzzy rules which maps the input features into output. These fuzzy rules are constructed using ‘if’ and ‘then’ statements for further classification [56, 57]. Sugeno FIS is computationally inexpensive compared to other defuzzification process viz. Mamdani FIS [56, 57].

14

ACCEPTED MANUSCRIPT

AdaBoost: AdaBoost is a meta-learning algorithm that combines result of various "weak" classifiers namely Least Square (LS), Normal Density Discriminant Function (NDDF), Perceptron, Pocket and Stumps adaptively to improve the classification performance [58]. The boosting algorithm iteratively calls the weak classifier in each run to feed different distribution of training data which improves the classifier

CR IP T

accuracy [58]. The performance of the aforementioned classifiers are validated using 10-fold cross validation and performance measures namely average accuracy, sensitivity, specificity and positive predictive value (PPV) are computed (please refer to Table

AN US

1). Table 1. Confusion matrix

TN- True Negative

FP- False Positive

M

Positive

FN- False Negative

Positive

ED

Negative

Negative

PT

(Classifier result)

Test outcome

Gold standard

TP- True Positive

CE

(8)

AC

(9) (10) (11)

TP: FLD class correctly diagnosed as FLD. FP: Normal class incorrectly identified as FLD. TN: Normal class correctly identified as Normal. FN: FLD class incorrectly identified as Normal. 15

ACCEPTED MANUSCRIPT

3.7 Fatty Liver Disease Index (FLDI) In this work, we have devised an integrated index named as FLDI using highly ranked LSDA components namely LSDA2 and LSDA3 (please refer to Figure 5 and Table A1). The concept of integrated index is proposed by Ghista [59] and Acharya et al. [60]. Later application of integrated index extended to detection of sudden

CR IP T

cardiac death [61], and identification of focal electroencephalogram (EEG) [62], depression [63], glaucoma [64], diabetic retinopathy [65] and oral cancer [66]. Equation (12) is formulated using trial and error to obtain FLDI which discriminates normal and FLD accurately using a single number.

4. Results

)

(

AN US

(

)

(12)

Total of 127,980 1D DCT coefficients are obtained from the ultrasound images. These

M

coefficients are reduced to 30 LSDA components using LSDA data reduction method. The reduced dimension of the LSDA components can be selected manually.

ED

In this work we have selected 30 reduced LSDA components. Further, reduced LSDA components are ranked using mRMR method to select best possible features

PT

to obtain highest classification performance to classify in to normal and FLD classes. Figure 5 shows the bar plot of mean values of ranked LSDA components for normal

CE

and FLD classes. It reveals that LSDA1, LSDA8 and LSDA3 have distinct values (please refer to Figure 5 and Figure 6) for normal and FLD class as compared to

AC

other LSDA components. Table A1 shows the summary statistics (mean, SD values, criterion value and rank) of LSDA components. The dimension of the LSDA components can be selected manually. However, in this work we have selected 30 LSDA components. The maximum classification performance is obtained using five LSDA components (please refer to Table 2 and Figure 5).

16

CR IP T

ACCEPTED MANUSCRIPT

AN US

Figure 5. Bar plot of mean values of ranked LSDA components for normal and FLD classes.

It can be seen from the Table A1 that the features with high criterion value obtained first rank and vice versa. These ranked features are fed sequentially to various supervised classifiers to identify the best possible features which maximize the

M

classification performance in discriminating normal and FLD classes. In this work, combination of first five ranked LSDA components yielded the highest classification

ED

performance. Moreover, 3D plot (please refer to Figure 6) shows that LSDA1, LSDA3 and LSDA8 have highest discrimination ability, since all these features have less

PT

overlapping area for normal and FLD classes. Hence, combination of these features obtained highest classification performance. Classification results in Table 2 shows

CE

that FS classifier obtained the highest average accuracy, sensitivity and specificity of 100% (please refer to Table 2 and Figure 7a) using first five ranked LSDA

AC

components (please refer to Table A1) compared to other classifiers (DT, k-NN, PNN and SVM).

17

CR IP T

ACCEPTED MANUSCRIPT

AN US

Figure 6: 3D plot of features LSDA1, LSDA3 and LSDA8 for normal and FLD classes.

Moreover, we have used ensemble learning algorithm viz. AdaBoost to classify normal and FLD classes using different weak learning methods. Since, the training and testing dataset is small (normal = 50 and FLD = 50). The results show that

M

AdaBoost with stumps weak learner obtained maximum average accuracy, sensitivity and specificity of 100% using 12 ranked LSDA components (please refer

ED

to Table 3 and Figure 7b). However, number of features used in AdaBoost is high

PT

compared to FS classifier to obtain the same maximum classification performance. Table 2: Performance measures of different classifiers.

NOF

CE

Classifiers

TP

TN

FP

FN

Accuracy (%)

PPV (%)

Sensitivity (%)

Specificity (%)

3

49

50

0

1

99

100

98

100

KNN (K=5)

2

49

48

2

1

97

96.08

98

96

PNN (σ=0.03)

6

50

49

1

0

99

98.04

100

98

SVM, RBF (σ=0.7)

6

48

50

0

2

98

100

96

100

SVM, Poly 1

6

50

48

2

0

98

96.15

100

96

SVM, Poly 2

3

48

49

1

2

97

97.96

96

98

SVM, Poly 3

3

48

49

1

2

97

97.96

96

98

Fuzzy Sugeno (σ=0.19)

5

50

50

0

0

100

100

100

100

AC

Decision Tree

[NOF: Number of Features]

18

ACCEPTED MANUSCRIPT

Table 3: Performance measures of AdaBoost classifier with different weak learning algorithms. AdaBoost with weak learners

TP

TN

FP

FN

Accuracy (%)

PPV (%)

Sensitivity (%)

Specificity (%)

LS

2

49

48

2

1

97

96.08

98

96

NDDF

29

48

50

0

2

98

100

96

100

Perceptron

7

49

49

1

1

98

98

98

98

Pocket

5

50

48

2

0

98

96.15

100

96

Stumps

12

50

50

0

0

100

100

100

100

(a)

AC

CE

PT

ED

M

AN US

[NOF: Number of Features]

CR IP T

NOF

(b) Figure 7. Plot of number of features vs. average performance using (a) Fuzzy sugeno classifier, and (b) AdaBoost classifier.

19

ACCEPTED MANUSCRIPT

The summary statistics (mean±SD) of range of FLDI is shown in Table 4 and Figure 8 reveals that FLDI values are distinct for normal and FLD classes. Table 4: Typical values of FLDI for normal and FLD classes. Normal 3.17±0.04

FLD 2.99±0.09

AN US

CR IP T

FLDI

Figure 8. Boxplot of range of FLDI for normal and FLD classes.

M

Moreover, in this work, we have compared the obtained results with existing feature

ED

extraction methods proposed in [29, 31, 36, 67] using ultrasound image dataset of current study. The most common feature extraction methods such as FOP [29, 31],

PT

GLCM [29, 31, 36], GLRLS [29], FD [29, 31], LTE [31], Bispectrum Entropies (BE) [36], DWT [36] and GIST [67] are used to discriminate normal and FLD classes (please

CE

refer to Table 6). The methodology used in current study obtained highest classification accuracy of 100% compared to other feature extraction methods

AC

mentioned in Table 5. Table 5: Comparison of best results of different feature extraction methods with current study for same dataset Method

Classifier

FOP, GLCM, GLRLS, FD FOP, GLCM,LTE, FD BE, GLCM, DWT GIST This study

NOF

Acc (%)

PPV (%)

Sen (%)

Spe (%)

SVM, Poly 1

19

85

84.31

86

84

SVM, Poly 1

33

89

89.80

88

90

Fuzzy Sugeno (σ=0.54) PNN (σ=0.11)

13 17

85 98

87.23 100

82 96

88 100

Fuzzy Sugeno (σ=0.19)

5

100

100

100

100

[NOF: Number of Features; Acc: Accuracy; Sen: Sensitivity; Spe: Specificity]

20

ACCEPTED MANUSCRIPT

5. Discussion It can be inferred from Table 6 that the texture feature methods (FD, GLCM, and wavelet) are used by many researchers to analyse FLD and normal liver US images. However, the Radon Transform and DCT based feature extraction method is not used in FLD detection so far. To the best of our knowledge this study is the first of its

CR IP T

kind using this method to FLD detection using ultrasound images. In our companion paper, we have extracted the GIST descriptors from the ultrasound images. Then these features are subjected to Marginal Fisher Analysis (MFA) integrated with Wilcoxon signed-rank test. The method used in this study is able to

AN US

diagnose the FLD with an average accuracy of 98%, sensitivity of 96% and specificity of 100% correctly using PNN classifier [67]. This current paper proposes a novel

automated FLD diagnosis algorithm using DCT coefficients or features obtained from the RT of images performed for every 1°angle up to 180°.The advantages of the

Automatically extracts the DCT frequency features from Radon transformed US images.



ED



M

methods used in this study are as follows.

Achieved an average accuracy, sensitivity and specificity of 100% using FS

PT

classifier in discriminating the two classes. FLDI (single number) is able to discriminate the two classes.



Developed algorithm is robust and repeatable as it uses ten-fold cross

CE



AC

validation technique. 

Used as an adjunct tool by the clinicians in their diagnosis.



Low frequency features are enhanced due to the RT, hence resulted in highest performance.

The limitations of our method is summarised below.

21

ACCEPTED MANUSCRIPT

Only 50 normal and 50 FLD images are used to evaluate the performance of our system. CAD systems required huge storage space and also may gradually affect the manual judging skills of the clinicians due to dependence on computers. Table 6: Summary of automated FLD detection methods using ultrasound images.

120 US images from 120 patients (60 used for training and 60 for testing)

First and second order gray level features, attenuation and backscattering features, backscattering coefficient

Pavlopoulos 150 US images et al., (2000) (normal + [29] cirrhosis + fatty)

ED

GLCM and nonseparable wavelet transform (NSWT) 20 liver US Textural features, Haar images from 10 wavelet detail energies patients (5 with FLD and 5 N)

CE

Ricardo et al., (2009) [32]

US images of 19 subjects

Texture analysis (FD, SGLDM, GLDS, RUNL, FOP) features

AC

Lupsor et al., (2011) [30]

Results/Findings Minimum distance classifier, sensitivity = 76.5%; Bayes quadratic classifier, sensitivity = 91.9% Voting k-NN classifier, sensitivity 100%

12

Accuracy (classification rate) = 82.67%

10

Accuracy = 90.5%

Not mentioned

Accuracy = 95%, Sensitivity = 100%

3

The ROC is better for the attenuation coefficient as compared to FO mean for the prediction of Steatosis (0.741 vs 0.652, p=0.001) moderate steatosis (0.791 vs 0.719, p=0.043).

PT

Yeh et al., (2004) [16]

Total No of features 9

CR IP T

Features/Methods

AN US

Data Used

M

Authors (Year) Yasser et al., (1996) [28]

526 subjects US Attenuation coefficient images and First order textural features

22

ACCEPTED MANUSCRIPT

177 US images from 36 patients (after pre-processing -131 ROIs for N and 131 ROI for FLD)

FOS, GLCM, LTE, FD

325

SVM, accuracy 79.77%

Ricardo et al., (2012) [32]

75 US images from 75 patients (35 FLD and 40 N)

Texture features

36

Accuracy = 93.54% Sensitivity = 95.83%

Singh et al., (2012) [34]

30 US images

Texture analysis methods (SGLCM, SFM, TEM, FPS and fractals)

Acharya et al., (2012a) [36]

58 FLD and 42 normal liver US images

Texture, wavelet and HOS features

Minhas et al, (2012) [37]

Dan et al (2013) [38]

Acharya et al (2015) [67]

AN US

6

Fisher’s linear discriminative analysis, Accuracy = 92% DT, accuracy = 93.3%

88 subjects US Texture analysis using images (30 wavelet packet FLD and 39 transform (WPT) normal, 19 heterogeneous)

61

SVM, accuracy = 95%

120 subjects US Maximum and images (N + minimum attenuation, FLD) maximum and minimum grey level, variance, skewness, kurtosis 180 ultrasound Texture features liver images(80 normal, 100 FLD)

9

SVM, accuracy = 87.78% Random Forest, accuracy = 90.84%

7

Accuracy 95% Sensitivity: 100%

Ultrasound images of 50 normal and 50 patients with FLD

17

PNN classifier Accuracy: 98%, Sensitivity: 96%, Specificity: 100%

PT

ED

M

3

CE

AC

Singh et al (2014) [35]

CR IP T

Andreia et al., (2012) [31]

GIST descriptors

23

ACCEPTED MANUSCRIPT Current method

Ultrasound images of 50 normal and 50 patients with FLD

RT and DCT coefficients

5

FS classifier, Accuracy = 100%, Sensitivity=100%, Specificity=100%

FLDI=100%

CR IP T

2

6. Conclusion

In this paper, an automated diagnosis based on RT and DCT coefficients is used

AN US

for the classification of normal and liver affected by fatty liver disease. The DCT coefficients are extracted from the Radon projections of liver ultrasound images and are used to train the classifier after subjected to LSDA. Among the obtained DCT

M

coefficients (features), highly discriminatory significant features are used to train

ED

and build classifiers. Using only two features, the FS classifier presented the highest accuracy, sensitivity and specificity of 100%. Moreover, FLDI is able to discriminate

PT

normal and FLD clearly using just two features. Since the technique is fully

CE

automated and highly user friendly, it can be easily used in clinical practice.

AC

Appendix

Table A1: Summary statistics of LSDA coefficients for normal and FLD classes. LSDA components Normal FLD Criterion Rank value Mean SD Mean SD LSDA4

-1.69E+02

1.64E+02

-9.35E+01

2.32E+02

0.4824

1

LSDA2

3.74E+01

2.72E+02

3.85E+01

3.76E+00

0.3461

2

LSDA1

-1.45E+00

5.33E+01

-5.02E+01

2.02E+02

0.1793

3

LSDA8

-1.99E+02

2.59E+02

-6.44E+02

1.22E+02

0.1565

4

LSDA3

3.14E+01

2.48E+02

6.83E+01

6.98E+01

0.1139

5

LSDA7

1.27E+02

3.33E+02

-5.58E+01

2.61E+01

0.1060

6

LSDA6

1.13E+02

2.02E+02

1.35E+02

1.44E+02

0.1024

7

24

ACCEPTED MANUSCRIPT

1.11E+02

3.11E+02

1.07E+02

1.13E+01

0.0937

8

LSDA16

2.35E+02

2.66E+02

2.13E+02

4.71E+01

0.0890

9

LSDA14

-2.18E+02

8.49E+01

-1.95E+02

2.09E+02

0.0852

10

LSDA9

5.19E+02

2.90E+02

5.05E+02

1.51E+01

0.0785

11

LSDA15

1.03E+01

2.67E+02

2.59E+00

3.48E+01

0.0776

12

LSDA10

-4.08E+02

3.15E+01

-3.78E+02

2.88E+02

0.0749

13

LSDA22

1.92E+01

3.82E+01

-6.98E+00

2.78E+02

0.0687

14

LSDA21

4.42E+02

2.55E+01

4.27E+02

2.43E+02

0.0449

15

LSDA17

-2.55E+02

2.20E+02

-2.38E+02

1.63E+01

0.0385

16

LSDA13

-2.74E+02

1.96E+01

-2.77E+02

2.49E+02

0.0335

17

LSDA11

2.68E+01

2.03E+02

3.06E+01

1.28E+02

0.0319

18

LSDA19

-7.60E+01

2.74E+02

-7.92E+01

8.87E+00

0.0239

19

LSDA12

-7.89E+02

1.93E+02

-7.89E+02

3.28E+02

0.0127

20

LSDA24

-7.64E+01

6.12E+01

-4.90E+01

2.97E+02

0.0127

21

LSDA23

7.38E+00

3.45E+02

1.22E+01

3.59E+01

0.0105

22

LSDA27

4.38E+01

2.16E+02

5.39E+01

7.34E+00

0.0094

23

LSDA29

-5.15E+01

2.49E+02

-4.24E+01

1.31E+01

0.0077

24

LSDA26

6.60E+01

2.11E+02

8.50E+01

1.16E+02

0.0063

25

LSDA18

-1.51E+02

1.78E+01

-1.50E+02

2.43E+02

0.0039

26

LSDA30

-1.06E+02

2.82E+01

-9.44E+01

2.98E+02

0.0021

27

LSDA20

1.66E+02

2.72E+02

1.66E+02

2.55E+01

0.0012

28

LSDA28

-1.55E+02

1.53E+01

-1.51E+02

2.45E+02

0.0005

29

LSDA25

-1.51E+02

1.14E+02

-1.47E+02

2.13E+02

0.0002

30

AN US

M

ED

References

CR IP T

LSDA5

PT

1. J. Ludwig, T.R. Viggiano, D.B. McGill, B.J. Oh. Nonalcoholic steatohepatitis: Mayo Clinic experiences with a hitherto unnamed disease. Mayo Clin Proc 55 (1980) 434-

CE

438.

2. G. Bedogni, L. Miglioli, F. Masutti, C. Tiribelli, G. Marchesini, S. Bellentani.

AC

Prevalence of and risk factors for non-alcoholic fatty liver disease: the Dionysos nutrition and liver study. Hepatology 42 (2005) 44-52.

3. J.D. Browning, L.S. SZczepaniak, R. Dobbins, P. Nuremberg, J.D. Horton, J.C. Cohen, S.M. Grundy, H.H. Hobbs. Prevalence of hepatic steatosis in an urban population in the United States: impact of ethnicity. Hepatology 40 (2004) 13871395.

25

ACCEPTED MANUSCRIPT

4. J.B. Schwimmer, R. Deutsch, T. Kahen, J.E. Lavine, C. Stanley, C. Behling et al., Prevalence of fatty liver in children and adolescents. Pediatrics 118 (2006) 1388-1393. 5. A. Duseja. Nonalcoholic fatty liver disease in India: A lot done, yet more required! Indian J Gastroenterol 29 (2010) 217-25. 6. J.G. Fan, G.C. Farrell. Epidemiology of non-alcoholic fatty liver disease in China. J Hepatol 50 (2009) 204-210.

CR IP T

7. G. Targher, L. Bertolini, R. Padovani et al. Prevalence of non-alcoholic fatty liver disease and its association with cardiovascular disease among type 2 diabetic patients. Diabetes care 30 (2007) 1212-8.

8. C.D. Williams, J. Stengel, M.I. Asike et al., Prevalence of non-alcoholic fatty liver

AN US

disease and non-alcoholic steatohepatitis among a largely middle-aged population utilizing ultrasound and liver biopsy: a prospective study. Gastroenterology 140 (2011) 124-31.

9. A.J. Sanyal. American gastroenterological association. AGA technical review on

M

non-alcoholic fatty liver disease. Gastroenterology 123 (2002) 1705-25. 10. D. Amarapurkar, P. Kamani, N. Patel, P. Gupte, P. Kumar, S. Agal, R. Baijal, S. Lala,

ED

D. Chaudhary, A. Deshpande. Prevalence of non-alcoholic fatty liver disease: population based study. Ann Hepatol 6 (2007) 161-3.

PT

11. S. Kojima, N. Watanable, M. Numata, T. Ogawa, S. Matsuzaki. Increase in the

CE

prevalence of fatty liver in Japan over the past 12 years: analysis of clinical background. J Gastroenterol 38 (2003) 954-61.

AC

12. J.G. Fan, Y.D. Peng. Metabolic syndrome and non-alcoholic fatty liver disease: Asian definitions and Asian studies. Hepatobiliary Pancreat Dis Int 6 (2007) 572-8.

13. I. Hasan, R.A. Gani, R. Machmud et al., Prevalence and risk factors for non-alcoholic fatty liver in Indonesia. J Gastroenterol Hepatol 17 (2002) S154. 14. S. Icer, A. Coskum, T. Ikizceli. Quantitative grading using grey relational analysis on ultrasonographic images of a fatty liver. Journal of medical systems 36 (4) (2012) 2521-2528.

26

ACCEPTED MANUSCRIPT

15. G. Li, Y. Luo, W. Deng, X. Xu, A. Liu, E. Song. Computer aided diagnosis of fatty liver ultrasonic images based on support vector machine, 30th annual international IEEE EMBS conference (2008). 16. W.C. Yeh, Y.M. Jeng, C.H. Li, P.H. Lee, P.C. Li. Liver fatty change classification using 25MHz high frequency ultrasound. IEEE ultrasonics symphosium 3 (2004) 2169-2172.

CR IP T

17. S.A. Harrison, S. Torgerson, P.H. Hayashi. The natural history of non-alcoholic fatty liver disease: a clinical histopathological study. Am J Gastroenterol 98 (2003) 20422047.

18. J. M. Clark, A.M. Diehl. Nonalcoholic fatty liver disease: an under recognised cause

AN US

of cryptogenic cirrhosis. JAMA 289 (2003) 3000-3004.

19. M.S. Ascha, I.A. Hanounch, R. Lopez, T.A.R. Tamini, A.F. Feldstein, N.N. Zein. The incidence and risk factors of hepatocellular carcinoma in patients with non-alcoholic steatohepatitis. Hepatology 51 (2010) 1972-1978.

M

20. E. Bugianesi, N. Leone, E. Vanni, G. Marchesini, F. Brunello, P. Carucci et al., Expanding the natural history of non-alcoholic steaohepatitis: from cryptogenic

ED

cirrhosis to hepatocellular carcinoma. Gatsroenterology 123 (2002) 134-140. 21. M. Charlton. Cirrhosis and liver failure in non-alcoholic fatty liver disease: molehill

PT

or mountain? Hepatology 47 (2008) 1431-1433.

CE

22. L.A. Adams, J.F. Lymp, J. St Sauver, S.O. Sanderson, K.D. Lindor, A. Feldstein et al., The natural history of non-alcoholic fatty liver disease: a population-based cohort

AC

study. Gastroenterology 129 (2005) 113-121. 23. M. Sasso, V. Miette, L. Sandrin. Novel controlled attenuation parameter for the evaluation of fatty liver disease. IEEE international ultrasonics symphosium proceedings (2009) 2256-2259. 24. J.W. Jeong, S. Lee, J.W. Lee, D.S. Yoo, S. Kim. Computer-assisted sonographic analysis of the hepatorenal and textural features for the diagnosis of the fatty liver. 27th annual conference engineering in medicine and biology (2005).

27

ACCEPTED MANUSCRIPT

25. L.A. Adams, J.A. Talwalkar. Diagnostic evaluation of non-alcoholic fatty liver disease. J clin Gastroenterolo 40 (2006) S34-S38. 26. A. Wieckowska, A.E. Feldstein. Diagnosis of non-alcoholic fatty liver disease: invasice versus noninvasive. Semin Liver Dis 28 (2008) 386-395. 27. U.R. Acharya, O. Faust, F. Molinari, S.V. Sree, S.P. Junnarkar, K.S. Vidya. Ultrasound-based tissue characterization and classification of fatty liver disease: A

CR IP T

screening and diagnostic paradigm, Knowledge-Based Systems 75 (2015) 66-77.

28. M.K. Yasser, A.F. Aly, M.Z. Jacek, M.B. Ahmed, Abou-Bakr MY. Classification algorithms for quantitative tissue characterisation of diffuse liver disease from ultrasound images. IEEE transaction on medical imaging 15 (1996) 466-477.

AN US

29. S. Pavlapoulos, E. Kyriacou, D. Koutsouris, K. Blekas, A. Stafylopatis, P. Zoumpoulis. Fuzzy neural network-based texture analysis of ultrasonic images, Improving accuracy of computer-assisted characterisation of diffuse liver diseases. IEEE engineering in medicine and biology (2000).

M

30. M. Lupsar, R. Badea, C. Vicas, S. Nedevschi, H. Stefanescu, M. Grigorescu, C. Radu, D. Crisan. Noninvasive steatosis assessment through the computerised processing

ED

of ultrasound images: Attenuation versus first order texture parameters. Meditech (2011).

PT

31. A. Andreia, S.S. Jose, S. Jaime, B.S. Pedro. Classifier approaches for liver steatosis

CE

using ultrasound images. Procedia technology 5 (2012) 763-770. 32. R. Ricardo, S. Joao. Fatty liver characterisation and classification by ultrasound.

AC

Pattern recognition and image analysis, Lecture notes in computer science 5524 (2009) 354-361.

33. R. Ricardo, T.M. Rui, J.M. Sanches. Global and local detection of liver steatosis from ultrasound. 34th annual international conference of the IEEE EMBS (2012). 34. M. Singh, S. Singh, S. Gupta. A new quantitative metric for liver classification from ultrasound images. International journal of computer and electrical engineering 4 (2012) 605-607.

28

ACCEPTED MANUSCRIPT

35. M. Singh, S. Singh, S. Gupta. An information fusion based method for liver classification using texture analysis of ultrasound images. Information Fusion 19 (2014) 91-96. 36. U.R. Acharya, S.V. Sree, R. Ribeiro, G. Krishnamurthi, R.T. Marinho, J. Sanches, J.S. Suri. Data mining framework for fatty liver disease classification in ultrasound: a hybrid feature extraction paradigm. Med Phys, 39 (2012a) 4255-64.

ultrasound images. J Med Syst 36 (2012) 3163-72.

CR IP T

37. F. Minhas, D. Sabih, M. Hussain. Automated classification of liver disorders using

38. M.M. Dan, G. Vasile, I.T. Corneliu, P. Alina, S. Ioan. Computer aided diagnosis method for steatosis rating in ultrasound images using random forests. Med

AN US

Ultrason 15 (2013) 184-190.

39. D.V. Jadhav, R.S. Holambe, Radon and discrete cosine transforms based feature extraction and dimensionality reduction approach for face recognition. Signal Processing 88 (2008) 2604-2609.

M

40. S. M. Pizer, E. P. Amburn, J. D. Austin, R. Cromartie, A. Geselowitz, T. Greer, B. ter Haar Romeny, J. B. Zimmerman, K. Zuiderveld. Adaptive histogram equalization

ED

and its variations, Computer vision, graphics, and image processing 39 (3) (1987) 355–368.

PT

41. E. Magli, G. Olmo, L.Lo. Presti. Pattern recognition by means of the Radon

CE

transform and continuous wavelet transform. Journal signal processing 73 (1999) 277-289.

AC

42. K. Jafari-Khouzani, H. Soltanian-Zadeh. Rotation-invariant multiresolution texture analysis using radon and wavelet transforms. IEEE Trans Image Process 14 (2005) 783-95.

43. N. Ahmed, T. Natarajan, K.R. Rao. Discrete cosine transform. IEEE Trans Computers 23 (1974) 90-93. 44. D. Cai, X. He, K. Zhou, J. Han, H. Bao. Locality sensitive discriminant analysis. In proceedings of the 20th international joint conference artificial intelligence (2007) 708-713. 29

ACCEPTED MANUSCRIPT

45. H. Peng, F. Long, C. Ding C. Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell 27(8) (2005) 1226–1238. 46. Z. Zhuo, K.C. Keong, L. Jiang, C.C.Y. Lui, A. Tin, W. Tien. Automatic glaucoma diagnosis with mRMR based feature selection. J Biom Biostat S7 (2012) 1–8. 47. R.N. Khushaba, K. Sarath, L. Sara, D. Gamini. Driver drowsiness classification using

CR IP T

fuzzy wavelet-packet-based feature extraction algorithm. IEEE Trans Biomed Eng 58(1) (2011) 121–131.

48. L. Breiman , J. Friedman, C.J. Stone, R. A. Olshen. Classification and regression trees. CRC press, 1984.

AN US

49. T.M. Cover, P.E. Hart. Nearest neighbor pattern classification. Information Theory, IEEE Transactions on 13(1) (1967) 21-27.

50. P.Cunningham, S.J. Delany. k-Nearest neighbour classifiers. Multiple Classifier Systems (2007) 1-17.

M

51. D.F. Specht, F. Donald F. Probabilistic neural networks. Neural networks 3(1) (1990) 109-118.

ED

52. L. Yang, S. Yang, S. Li, R. Zhang, F. Liu, L. Jiao. Coupled compressed sensing inspired sparse spatial-spectral LSSVM for hyperspectral image classification.

PT

Knowledge-Based Systems 79 (2015) 80-89.

CE

53. D. Tomar, S. Agarwal. A comparison on multi-class classification methods based on least squares twin support vector machine. Knowledge-Based Systems 81 (2015)

AC

131-147.

54. J. Han, M. Kamber, J. Pei. Data mining: concepts and techniques. Waltham, MA: Morgan Kaufmann (2005).

55. Y.S. Jeong, R. Jayaraman. Support vector-based algorithms with weighted dynamic time warping kernel function for time series classification. Knowledge-Based Systems 75 (2015) 184-191. 56. Sugeno, Michio. Industrial applications of fuzzy control. Elsevier Science Inc., 1985.

30

ACCEPTED MANUSCRIPT

57. Ross, Timothy J. Fuzzy logic with engineering applications. John Wiley & Sons, 2009. 58. Y. Freund, R. E. Schapire, A decision-theoretic generalization of on-line learning and an application to boosting, Journal of Computer and System Sciences 55 (1) (1997) 119 – 139. 59. D. N. Ghista. Nondimensional physiological indices for medical assessment. Journal

CR IP T

of Mechanics in Medicine and Biology 9 (4) (2009) 643-669.

60. U. R. Acharya, O. Faust, S. V. Sree, D.N. Ghista, S. Dua, P. Joseph, V.I.T. Ahamed, N. Janarthanan, and T. Tamura. An integrated diabetic index using heart rate variability signal features for diagnosis of diabetes. Computer methods in

AN US

biomechanics and biomedical engineering 16 (2) (2013) 222-234.

61. U. R. Acharya, U. H. Fujita, V. K. Sudarshan, V. S. Sree, L.W.J. Eugene, D.N. Ghista, R. S. Tan. An integrated index for detection of Sudden Cardiac Death using Discrete Wavelet Transform and nonlinear features. Knowledge-Based Systems 83 (2015)

M

149-158.

62. R. Sharma, R. B. Pachori, and U. R. Acharya. An Integrated Index for the

ED

Identification of Focal Electroencephalogram Signals Using Discrete Wavelet Transform and Entropy Measures. Entropy 17 (8) (2015) 5218-5240.

PT

63. U.R. Acharya, V.K. Sudarshan, H. Adeli, J. Santhosh, J.E.W. Koh, S. D. Puthankatti,

CE

A. Adeli. A Novel Depression Diagnosis Index Using Nonlinear Features in EEG Signals. European neurology 74 (1-2) (2015) 79-83.

AC

64. M.R.K. Mookiah, U. R. Acharya, C. M. Lim, A. Petznick, and J. S. Suri. Data mining technique for automated diagnosis of glaucoma using higher order spectra and wavelet energy features. Knowledge-Based Systems 33 (2012) 73-82.

65. U.R. Acharya, E. Y. K. Ng, J. H. Tan, S. V. Sree, and K. H. Ng. An integrated index for the identification of diabetic retinopathy stages using texture parameters. Journal of medical systems 36 (3) (2012) 2011-2020. 66. M.R.K. Mookiah, V. Venkatraghavan, U. R. Acharya, M. Pal, R. R. Paul, C. M. Lim, A. K. Ray, J. Chatterjee, C. Chakraborty. Automated oral cancer identification using 31

ACCEPTED MANUSCRIPT

histopathological images: a hybrid feature extraction paradigm. Micron 43 (2) (2012) 352-364. 67. U.R. Acharya, H. Fujita, S. Bhat, U. Raghaventra, A. Gudigar, F. Molinari, A.Vijayanathan, K.H. Ng. Decision support system for fatty liver disease using GIST descriptors extracted from ultrasound images. Information Fusion 29 (2016)

CR IP T

32--39.

AC

CE

PT

ED

M

AN US

Graphical Abstract

32