PERFORMANCE ANALYSIS OF GRAY LEVEL CO- OCCURRENCE MATRIX TEXTURE FEATURES FOR GLAUCOMA DIAGNOSIS

Glaucoma is a multifactorial optic neuropathy disease characterized by elevated Intra Ocular Pressure (IOP). As the visual loss caused by the disease is irreversible, early detection is essential. Fundus images are used as input and it is preprocessed using histogram equalization. First order features from histogram and second order features from Gray Level Co-occurrence Matrix (GLCM) are extracted from the preprocessed image as textural features reflects physiological changes in the fundus images. Second order textural features are extracted for different quantization levels namely 8, 16, 32, 64, 128 and 256 in four orientations viz 0, 45, 90 and 135° for various distances. Extracted features are selected using Sequential Forward Floating Selection (SFFS) technique.The selected features are fed to Back Propagation Network (BPN) for classification as normal and abnormal images. The proposed computer aided diagnostic system achieved 96% sensitivity, 94% specificity, 95% accuracy and can be used for screening purposes. In this study, the analysis of gray levels have shown their significance in the classification of glaucoma.


INTRODUCTION
Glaucoma, the second leading cause of blindness is characterized by the progressive degeneration of optic nerve fibers, the "cable" that transmits the visual message from the eye to the brain and characteristic damage to the visual field. A survey shows that the glaucoma is responsible for 6 million cases worldwide. The main risk factor of glaucoma is elevated pressure in the eye which causes damage to the eye optic nerve. It also affects the optic disc by enlarging the cup size. It is possible that early functional changes are detectable before significant loss occurs, at a stage where they may still be reversible through early intervention.
As last capabilities of the optic nerve cannot be recovered, early diagnosis is important to stop or slow down disease progression. Presently, state-of-the art glaucoma detection requires mass screening. A glaucoma diagnosis system based on texture features using the less expensive digital fundus images are analyzed to discriminate normal and glaucomatous images.
Glaucoma Risk Index (GRI) was described by Bock et al. (2010) to assess glaucoma. The accuracy of 80% has been achieved in a 5-fold cross-validation setup and the GRI gains an Area Under Convergence (AUC) of 88%. Tan et al. (2010) analyzed ocular thermal images with second-order texture features together with first-order texture analysis, moments and difference histogram. Several significant linear correlations among investigated features were observed. The features extracted from cross co-occurrence matrix played an important role in the diagnosis of ocular diseases.
Texture provides important information about the structural arrangement of surfaces and their relationship to the surrounding environment. The textural properties of images appear to carry useful information for discrimination purposes and hence texture analysis of fundus image is performed. Ramamurthy and Chandran (2012)

developed
Science Publications AJAS an efficient image retrieval systems using Gray Level Cooccurrence Matrix statistical measures computed provide increasing similarities between query image and database images for improving the retrieval performance along with the large scalability of the databases. Veluchamy et al. (2012) segmented blood cells using morphological operations and extracted first, second order gray level statistics and geometrical parameters. Features extracted achieved 80% classification accuracy. Kavitha and Duraiswamy (2011) used texture features to discriminate between normal and nonproliferative eyes from the quantitative assessment of monocular fundus images. Two parameters energy and contrast were considered to be the most efficient for discriminating different textural patterns. A novel glaucoma detection system using a combination of texture and higher order spectra features proposed by Rajendra et al. (2011) provided an accuracy of more than 91%. Yogesh and Sasikala (2012) described texture analysis of retinal layers in spectral domain OCT in which diagnosis of age related macular degeneration, diabetic macular edema was tested. The behavior of cooccurrence statistics was investigated by Gomez et al. (2012) to classify Breast Ultrasound (BUS) images. The best AUC of 0.81 was achieved with 32 gray levels and 109 features. AUC of 0.87 was obtained for single texture feature. The texture descriptors that contributed notably to distinguish lesions were contrast and correlation. Krishnan and Oliver (2013) proposed an automated glaucoma diagnosis system based on hybrid feature extraction from digital fundus images. Further a novel integrated index called Glaucoma Risk Index (GRI) which is composed from HOS, TT and DWT features, is used to diagnose the unknown class using a single feature during the mass screening of normal/glaucoma images. Most of the shape based approaches assume a valid segmentation of the optic disk. However, a small error in these segmentation based techniques may result in significant changes in the measurements and leads to errors in the diagnosis. As feature extraction is performed in the digital domain, the proposed glaucoma diagnosis system can be easily incorporated into existing medical image analysis for large scale screening. So a glaucoma diagnosis system is proposed based on the analysis of gray levels in different orientations using digital fundus images (Fig. 1). Nyul (2009) focused on a novel automated classification system for glaucoma, based on image features from fundus photographs. Extracted high-dimensional feature vectors are compressed via Principal Component Analysis (PCA) and combined with Support Vector Machine (SVM). In this technique, accuracy of detecting glaucomatous retina fundus images are comparable to that of human experts.

AJAS
Texture features are chosen since fundus images reflect the physiological changes that occur due to glaucoma progression. Glaucoma diagnosis system is proposed in this study based on the behavior of texture features as a function of gray-level quantization to classify retinal images. Different gray levels are analyzed to clearly discriminate between the normal and abnormal stages.

MATERIALS AND METHODS
After preprocessing, the acquired fundus images, texture based features are extracted for different quantization levels and subjected to feature selection process. Subsequently, these features are fed to back propagation network for classification. Further an analysis on the influence of different features is performed to assess glaucoma.

Data
Fundus images are captured by Topcon TRC50 EX mydriatic fundus camera with a 50° field of view at Aravind eye hospital, Madurai. The image size is 1900×1600 pixels at 24 bits true color image. Doctors in the ophthalmic department of the hospital approved the images for the research purpose. Various stages of glaucoma affected fundus images are taken from http://www.optic-disc.org, a public database.

Preprocessing
The retinal images are pre-processed to improve the local contrast of an image and bring out more detail.
Image contrast is enhanced using histogram equalization by transforming the values in an intensity image. Figure  2 describes the histogram equalized image.

Texture Features
Texture analysis refers to the description of characteristic image properties by textural features. Texture is a result of local variations in brightness within one small region of an image. Naseri et al.
(2012) described texture as a measure of surface roughness when the intensity values of an image are considered to be elevations.
Four first order features namely Mean, Standard deviation, Entropy and Variance are extracted from the preprocessed image. Second order textural features are then extracted using Gray-Level Co-occurrence Matrix (GLCM) to represent a set of features to reduce the misclassification of glaucomatous images. GLCM depicts how different combinations of pixel brightness values occur in an image.It is a secondorder measure because it measures the relationship between neighboring pixels. GLCM represents the joint frequencies of all pair wise combinations of gray levels i and j separated by a distance along a direction θ as shown in Equation (1). Then, the GLCM can be defined as: (1)

Fig. 2. Grayscale fundus image and histogram equalized image
Science Publications

AJAS
Where: (x 1 , y 1 ) and (x 2 , y 2 ) = Pixels in the Region Of Interest (ROI) I ( . ) = The gray-level of the pixel || . || = The number of the pixel pairs that satisfy the conditions Each GLCM was normalized by the sum of all its elements to calculate the co-occurrence relative frequency between gray level i and j. Each entry in GLCM matrix is considered to be the probability that a pixel with value i will be found adjacent to a pixel of value j as in Equation (2): Twenty two features were extracted from the GLCM matrix viz. autocorrelation, contrast, correlation I, correlation II, cluster prominence, cluster shade, dissimilarity, energy, entropy, homogeneity I, homogeneity II, maximum probability, sum of squares, sum average, sum entropy, sum variance, difference variance, difference entropy, information measure of correlation I, information measure of correlation II, inverse difference normalized and inverse difference moment normalized are extracted from the GLCM matrix. The process is continued for six quantization levels (8, 16, 32, 64, 128 and 256) for orientations of 0, 45, 90 and 135° in various distances d = 1,2,3,4. Thus twenty four co-occurrence matrices were obtained for graylevels in different orientations.

Feature Selection
High dimension data could contain high degree of redundant information and degrade the efficiency of the system. Hence feature selection process is performed using Sequential Floating Forward Selection (SFFS) feature selection method. This method is used to deal with the nesting problem. The best subset of features was initialized as the empty set and at each step a new feature was added. After that, the algorithm searches for features that can be removed until the correct classification error does not increase. The "best subset" of features is constructed based on the frequency with which each attribute is selected. The selected features are provided as input for the classification.

Classification
Back propagation is a common method of training artificial neural networks. The network learns from many inputs, to get the desired output. It is a supervised learning method and is a generalization of the delta rule. It requires a dataset of the desired output for many inputs, making up the training set. It is most useful for feed-forward networks (networks that have no feedback, or simply, that have no connections that loop). Helmy and El-Taweel (2010) developed a neural network model for satellite images based on both spectral and textural analysis. Better discrimination with 23% higher accuracy was achieved in the trained network with textural features than without textural parameters.

Performance Analysis of Textural Features
Discrimination of normal and abnormal images for different textural features are analyzed using box-plot. The box-plot for three features namely contrast, entropy and cluster prominence in four orientations 0, 45, 90 and 135° are shown from Fig. 3-5.
From the GLCM matrix, second order features were extracted in different orientations for gray levels of 8, 16, 32, 64, 128. Feature values derived for gray levels 8 and 16 are shown in Table 1 and 2 has the feature values derived for gray level 32 and 64 for an orientation of 0°.  Feature values for gray level 32 analyzed for normal and abnormal images in orientations (0°, 45°) are shown in Table 3 and orientations (90°, 135°) are shown in Table 4.

RESULTS
Performance of the classifier are measured in terms of sensitivity, specificity and accuracy. Sensitivity is a measure that determines the probability of results that are true positive such that the person is affected by the disease. Specificity is a measure that determines the true negatives such that the person is not affected by retinal disease. Accuracy is a measure that determines the results that are accurately classified. MATLAB version 7.12 is used for implementation of the work. Comparative analysis for all gray level quantization namely 8, 16, 32, 64, 128 and 256 are performed based on correctly classified images.
Impact of individual gray levels on the detection of normal and abnormal fundus image are given in Table 5.
Graph showing the performance evaluation of various gray level quantization are shown in Fig. 6.

Training Performance Using BPN
The back propagation neural network was used for the classification of the fundus images. Classification is performed for all gray level quantization to determine the best discriminating gray level and the result shows that the gray level 32 classifies with higher accuracy. In this study, 50 images are used for training and 100 images for testing. About 50 images (25 normal and 25 abnormal) for training and 100 images (50 normal and 50 abnormal) for testing were used for classification. From the available dataset, data is split into set1 and testing set. Next, set1 is further divided into training and validation set. Then the classifier is trained using training set and tested on validation set. The process is repeated by selecting various combinations of training and validation set. The classifier which gives best performance is then selected and used to get performance in the testing set. The sample neural network results containing Receiver Operating Characteristic (ROC) curve plotted against false positive and true positive rate are shown in Fig. 7.

DISCUSSION
Quantitative texture analysis was performed to improve diagnosis of retinal images. Features derived from the 32 gray levels classified 47 normal images correctly out of 50 normal images and 48 abnormal images as abnormal out of 50 abnormal images with an accuracy of 95%. The accuracies obtained for gray levels 8 and 16 were 87% and 89% respectively. For the gray level 256, the accuracy was about 89% and the computation complexity also increases. The gray levels 16 and 128 produced accuracy of 89% and 91% respectively, but its specificity was greater than the sensitivity. It means that the abnormal images were classified as normal which was undesirable. The gray level 32 classified the fundus images with specificity of 94%, sensitivity of 96% and accuracy of 95%. Thus gray level 32 holds good for the classification of the fundus images. First order feature viz. skewness and second order features contrast, correlation, dissimilarity, cluster prominence and autocorrelation showed good discrimination of normal and abnormal retinal images at 90° orientation for unit distance.

CONCLUSION
This study presented a study of co-occurrence texture statistics as a function of gray-level quantization for glaucoma diagnosis. The retinal images were preprocessed for improving the local contrast of an image and bringing out more detail. Twenty-two texture features autocorrelation, contrast, correlation I, correlation II, cluster prominence, cluster shade, dissimilarity, energy, entropy, homogeneity I, homogeneity II, maximum probability, sum of squares, sum average, sum entropy, sum variance, difference variance, difference entropy, information measure of correlation I, information measure of correlation II, inverse difference normalized and inverse difference moment normalized extracted from the GLCMs were extracted for different quantization levels 8, 16, 32, 64, 128 and 256 for 0, 45, 90 and 135° orientations for unit distance. Sequential Forward Floating Selection (SFFS) was used for feature selection. The selected features were fed as input to back propagation network to classify Science Publications AJAS the images as normal and abnormal. Features computed from GLCM with orientation of 90°, quantization level 32, unit distance contributed notably to distinguish between normal and abnormal fundus images. The proposed computer aided diagnostic system achieved 96% sensitivity, 94% specificity, 95% accuracy and can be used for screening purposes.
When more number of gray levels is included in GLCM, computational cost of texture statistics becomes higher. GLCM statistics can be used to discriminate between normal and abnormal images but GLCM texture features performs poorer in differentiating stages of abnormalities. Too many texture features affects classification rate and hence feature selection or significant tests are highly essential to identify selective features. An efficient automatic system can be developed by including diverse images for the mass screening of glaucoma. Classification accuracy can be improved by increasing the number of training images and choosing better features. To identify disease progression, GLCM statistics can be combined with other texture features or geometrical features.