Automatic Diagnosis of Glaucoma Using Digital Fundus Images

One application of GMM is to form smooth approximations for arbitrarily shaped densities. GMM provides a useful tool to model the characteristics of multi-model distributed data. Another useful characteristic is the fact that GMM employs a diagonal covariance matrix that is less complex compared to full covariance matrixes that are usually required.18 This significantly reduces the computational complexity of the algorithm. GMM has been used in many areas such as pattern recognition and classification. In general, this method poses a great success in the areas of identification and verification.

6.3. Results

We start the result discussion by listing both mean and standard deviation of the computed features. Table 6.1 shows this list. The c/d ratio and the area of blood vessels are larger for glaucoma, due to the increase in pressure. This ratio is 0.343 ± 0.245 for a normal subject and 0.503 ± 0.221 for a glaucoma subject. The number of blood vessels for a normal subject is 29254.3 ± 10775.5. This number increases in glaucoma subjects (35746 ± 11443.2). The ISNT ratio is also greater for subjects suffering from glaucoma (1.037±0.021) than the normal subjects (1.024±0.02). We conducted a Student t-test on these two groups (normal subjects and glaucoma subjects) for different features, and the acquired p value was less than 0.03. The low p value indicates that these results are statistically significant.

Table 6.2 illustrates how many samples were used for training and testing the classifier. Furthermore, this table lists also the classification results. In this investigation, 42 images were used for training and the rest (18 images) were used for testing. During classification, only one normal sample was

Table 6.1. Values of three features for normal and glaucoma cases.

221

Rajendra Acharya U. et al.

Table 6.2. Classification results.

Table 6.3. Sensitivity, specificity, and positive predictive values for the GMM classifier.

classified as abnormal and two glaucoma images were classified as normal. The average classification rate was 83.33%.

Table 6.3 shows the sensitivity, specificity, and positive predictive accuracy for the two classes. In the table, we denote true positive (TP) for the number of glaucoma images classified as correctly as glaucoma, true negative (TN) for the number of normal images correctly identified as normal, false negative (FN) for the number of glaucoma samples misclassified as normal, and false positive (FP) for the number of normal image misclassified as glaucoma. Sensitivity is the probability of an abnormal subject being correctly classified as abnormal; specificity is the probability of a normal subject being correctly identified as normal by classifier. The proposed system detects glaucoma with a sensitivity of 77.78% and a specificity of 88.89%. Furthermore, the positive predictive value is 87.50%.

A graphical user interface (GUI) was developed in this work to enable a user to access the algorithm in a user-friendly manner, as illustrated in Fig. 6.12. It comprised input image data, output feature extraction data, a review of the patients’ last visit, patient data selection buttons and display,

222

Automatic Diagnosis of Glaucoma Using Digital Fundus Images

Fig. 6.12. The GUI.

and a classification textbox. Using the Patient Data button, the patient image file (Image 2) was loaded. Then, the original image, optic disc, optic cup, blood vessels, inferior blood vessels, superior blood vessels, nasal blood vessels, and temporal blood vessels were displayed. The patient details were also be displayed in the Patient Data section of the display. Patient ID, gender, age, attending physician, date of birth, race, date of scan, and previous visit were automatically displayed in the Patient Data section.

In addition, the right-hand corner of the display showed an earlier visit image. There is a provision provided to display optic disc and blood vessel images and the ISNT ratio.

6.4. Discussion

We have extracted three features to detect glaucoma automatically. Our features are clinically significant and can identify the disease with an accuracy of 83%. It is important to diagnose glaucoma in an early stage in order to minimize damage to the optic nerve. Only if this damage is minimal, the disease can be effectively treated and the progression of the disease can be prevented.

223

Rajendra Acharya U. et al.

Previously, six fuzzy classification algorithms were employed to detect the presence and the absence of glaucoma with a classification rate of less than 76%.10 An ANN was proposed to recognize glaucomatous visual field defects, and its diagnostic accuracy was compared with that of other algorithms.19 For this work, the Glaucoma Hemifield Test attained a sensitivity of 92% at 91% specificity. The ANN method itself achieved a sensitivity of 93% and a specificity of 94%. The area under the ROC curve was 0.984.

Bowd et al. used neural network techniques to differentiate glaucomatous and nonglaucomatous eyes. They extracted optic disc topography parameters from the Heidelberg retina tomograph.20 The areas under the ROC curves for SVM linear and SVM Gaussian were 0.938 and 0.945, respectively, for the MLP, the ROC area was 0.941, and for the LDF, the ROC area was 0.906. With the use of forward selection and backward elimination optimization techniques, the areas under the ROC curves for SVM Gaussian and the current LDF were increased to approximately 0.96. Hence, the neural network analyses show an increasing diagnostic accuracy of tests for glaucoma.

Recently, Nayak et al. have used a novel method for glaucoma detection using the c/d ratio, the ratio of the distance between optic disc center, the optic nerve head to diameter of the optic disc, and the ratio of blood vessels area in inferior-superior quadrants to the area of blood vessel in the nasal-temporal quadrants.14 The resulting feature vector was fed to a neural network for classification. Their proposed system classified the glaucoma automatically with a sensitivity and specificity of 100% and 80%, respectively.

Our results show a sensitivity of 77.7% and a specificity of 88.8% for the proposed system. We predict that the accuracy of the proposed classification system can be improved by using more parameters, such as textures. In addition, by increasing the number of training and testing images, the result can be further improved. The environmental lighting condition plays an important role in the determination of the classifier performance. Uniform lighting condition set while acquiring a fundus image can also yield better results. This method can serve as an adjunct tool to aid a physician in crosschecking his or her diagnosis.

224