Classification of Ultrasound Kidney Images using PCA and Neural Networks

In this paper, a computer-aided system is proposed for automatic classification of Ultrasound Kidney diseases. Images of five classes: Normal, Cyst, Stone, Tumor and Failure were considered. A set of statistical features and another set of multi-scale wavelet-based features were extracted from the region of interest (ROI) of each image and the principal component analysis was performed to reduce the number of features. The selected features were utilized in the design and training of a neural network classifier. A correct classification rate of 97% has been obtained using the multi-scale waveletbased features. Keywords—Ultrasound kidney images; Feature Extraction; Principal Component Analysis; Neural Network classifier


I. INTRODUCTION
In recent years, great advances have been made in automated systems for detecting kidney diseases using ultrasonic systems which allow a greater amount and quality of information to be extracted during imaging the patients.The use of feature extraction, image analysis and pattern recognition techniques for classification is most suited to the evaluation of global conditions (e.g.failure, stone, tumor, and cyst).
Previous work have utilized feature extraction techniques such as Gray Level Statistical features, Fourier Transform and Gabor Wavelet's features with kidney and liver ultrasound images [1][2][3].
The present paper describes an automated system for analyzing and classifying ultrasound kidney images.The system starts with capturing the ultrasound kidney image and identifying the region of interest.Image Preprocessing techniques are also employed to improve image quality and reduce noise.Discrete Wavelet Transform (DWT) was used for feature extraction as it has potential capacity in classification problems [4].Moreover, statistical features were extracted for the comparison purposes.Feature extraction with DWT yields to a large number of features being extracted, so the PCA technique was employed as being efficient in selecting the optimal features [1], [5], and [6].The last stage in the system is the classifier.A multi-layer neural network was designed and trained using the optimal features selected by PCA.
The characteristics of the designed classifier were investigated and optimized for both best performance and highest classification rate [7].Fig. 1 shows a block diagram for the proposed system.

A. Region-Of-Interest (ROI)
The first step in image preprocessing is to determine the region of interest (ROI).It will improve the speed and accuracy of classification process by selecting only the kidney and removing unneeded details like patient and scan information.Previous researches have proposed automated ways to get the ROI [9] nevertheless, in the present work, a rectangular ROI of size 256x256 was obtained manually by cropping the kidney image to simplify the process and limit the possibilities of errors.The ROI size of 256x256 pixels was chosen as being suitable for both longitudinal and transverse kidney images.Fig .3 shows a normal kidney image with outlined ROI area.

B. Speckle noise
To improve the quality of ultrasound kidney images, image-preprocessing techniques have been adopted.Three filters were applied and their performance was compared and evaluated in terms of entropy.These are: Wiener, Histogram Equalization and Median Filter.

1) Wiener Filter
Wiener filter is used to reduce the noise present in the image [10].Fig. 4 shows the output image after using Wiener filter.

2) Histogram Equalization
Histogram equalization is used to improve the visual appearance of an image by adjusting the image histogram [11], [12] .Fig. 5 shows the results of applying histogram equalization.

3) Median Filter
Median Filter helps in reducing mainly speckle and salt and pepper noise [12].Fig. 6 shows the resulting image after using a 3 x3 median filter.
The results showed that Median filter gives the best performance especially if the evaluator is concerning more on the kidney edges than the whole image.This is in agreement with that reported in [11] IV. FEATURE EXTRACTION

A. Statistical Features
Statistical features is one of the early methods proposed in image processing.The gray level co-occurrence matrix (GLCM) of the ROI was used as suggested by Haralick [13].
The following features are extracted from the GLCM of the ROI kidney images using MATLAB: Energy, Entropy, Contrast, Homogeneity, Maximum probability and correlation [14].
1) Energy is a measure of local homogeneity and it is calculated using: where, i and j are the pixel values.

B. Discrete wavelet Transform
The discrete wavelet transform (DWT) is a multi-resolution analysis technique that analyzes the signal by decomposing the signal into its coarse and detail information, this is accomplished by using successive high-pass and low-pass filtering operations [15], [16], based on following equations: , -∑ ( ) ( )

V. FEATURE SELECTION
The Principal Component Analysis (PCA) technique was used frequently in previous work for feature reduction in classification problems with ultrasound images [1], [5].PCA was also used in conjunction with DWT in other object classification problems like face recognition [17].In this work, PCA helps in reducing the feature vector dimension obtained from DWT of ultrasound kidney images.

Procedure for making PCA:
 Getting the covariance matrix.

 Getting the Eigen Vectors
 Selecting top Eigen Values from Eigen Vectors www.ijacsa.thesai.orgUsing MATLAB, the first step in this procedure results in a covariance matrix of size 32x32.Second step gives an eigenvector of 32 values i.e. the eigenvalues or the principal components .
The results have shown that the first 15 eigenvalues in the eigenvector, can be considered as useful-nonzero values.These first 15 values will then be used in the classification stage.

A. Neural Network Topology Design
Using "NeuralBuilder" module in "NeuroSolutions" software provided by "NeuroDimension", Inc., a multilayer neural network as shown in Fig. 8, with two hidden layers with 10 nodes each was designed.Five output nodes were used to produce the following output encoding for the five kidney image classes-"10000" for "Normal", "01000" for "Failure", "00100" for "Stone", "00010" for "Tumor" and "00001" for "Cyst".A Sanger´s rule and sigmoidal activation function were found suitable for the classification purpose.A mean square error value of 0.05 was used to stop the learning process.

B. Building and Training ANN
Network characteristics like number of hidden layers, processing elements in each layers, optimization method and learning rule are customizable and could be adjusted for getting better learning rate and less mean square error (MSE) which is an important measure of network performance.Different topologies of neural networks were used to reach the best results.
Optimal results were obtained when the number of processing elements in the first hidden layer is 3 nodes and in the second hidden layer is 6 nodes.Training was performed using the hold-out method where 50% of the data were used for training the classifier and 50% for testing [18].Each set of features was used separately and the results were compared.
A correct classification rate of 95% was obtained using statistical features; on the other hand, 97% of the images were correctly classified when trained with wavelet features.Table III illustrates the results for each feature set; where sensitivity and specificity were calculated using equations ( 9), ( 10) respectively.

Sensitivity= true positives/(true positive + false negative) (9)
Specificity= true negatives /(true negative + false positives) (10) VII.CONCLUSION In this paper, an automatic system for the detection and classification of kidney diseases has been developed.The system consists of five main parts: ROI segmentation, image preprocessing, feature extraction, feature selection and classification.ROI segmentation was performed manually with the help of the physician by cropping.Image preprocessing was carried out using three types of filters: Wiener filter, Median filter and Histogram Equalization filter.The results showed that Median filter gives the best performance.Two sets of features were extracted using two different features extraction techniques.These are statistical-based features and the multi-scale wavelet-based features.Feature selection was achieved using the principal component analysis approach.A multilayer feed forward neural network utilizing the backpropagation algorithm was used for classification purpose.It has been shown that the highest classification rate was obtained using the multi-scale wavelet-based features.A correct classification rate of 97% has been obtained which is comparable to similar neural networks classifiers used in [1], [2], [19].The results are encouraging and promising.Further work is required to apply the suggested methodologies to a larger data set with a wide spectrum of kidneys disorders and to develop a complete intelligent system that can be used as an assistant tool in automatic classification of ultrasound kidney images.Improving the classification accuracy is a subject of a current investigation which aims to develop a complete automatic kidney images classification.

Fig. 1 .
Fig. 1.The Proposed system for automatic classification of kidneys images The paper is organized as follows: Section II describes the database used and the kidney diseases considered in the present study.Image preprocessing techniques are explained in Section III.Sections IV and V present the feature extraction and feature selection techniques utilized, respectively.The classification stage is explained in Section VI where an artificial neural network is utilized.Section VII gives the concluding remarks.

5 )
within a group.Correlation is a measure of the strongest of the relationship between two variables.the results of feature extraction associated with each class of images.Values in the upper row are the mean values for each class and the lower row gives the standard deviation for each class.

TABLE I .
TYPES OF DISEASES AND THE NUMBER OF IMAGES USED FOR EACH TYPE

TABLE III .
COMPARISON OF THE CLASSIFICATION RESULTS