ICBME2008 Development and Verification of Artificial Neural Network Classifiers for Eye Diseases Diagnosis Hossein Parsaei 1* , Mohammad.H.Moradi 2 , Roya Parsaei 3 1- Systems Design Engineering Department, University of Waterloo, Waterloo, ON, Canada. 2-Biomedical Engineering School, Amir Kabir University of Technology, Tehran, Iran. 3-Mechanical Engineering School, Khajeh Nasir Toosi University of Technology, Tehran, Iran. * Email: hparsaei@engmail.uwaterloo.ca Abstract: Visual field sensitivity test results are crucial for accurate and efficient diagnosis of blinding diseases such as glaucoma, scotoma, homonymous, lesions of the optic nerves, lesions of the chiasm, etc. Typically in computerized perimeters, analysis of visual field sensitivity test results is performed by statistical methods. The purpose of these analyses is to help ascertain whether the test results are acceptable or not and also what the disorder is. Herein, first, Kohonen’s self – organizing map (SOM) is used to establish whether the Perimetry result is reliable or not. Then, Multilayer Perceptron (MLP), Probability Neural Network (PNN), Radial-Basis Function Network (RBFN) and Support Vector Machines (SVM) is used to analyze premetry results. By comparing the statistics, the artificial neural network classifiers show encouraging performance and SVM has the best performance. Keywords: Visual field, Standard automated perimetery, perimeter, Multi Layer Perceptron (MLP), Probability Neural Network (PNN), Radial-Basis Function Network (RBFN), Support Vector Machine (SVM). I. Introduction The visual field has been defined by Tate and Lynn (1977) as all the space that one eye can see at any given instant time [1]. It is normally measured in degrees from the line of sight. When a patient is looking straight ahead, his/her visual field on average extends 65 degrees upwards, 75 degrees down wards, 60 degrees inwards (towards the nose), and 95 degrees outwards (temporally) [1]. Visual field testing provides the eye-care practitioner with essential information regarding the early detection of major blindness-causing diseases such as glaucoma, scotoma and many other ocular and neurological diseases [2-6]. Testing the visual field, known as “perimetry,” is normally performed using a specially designed instrument called “perimeter”. Perimetry requires a medical operator to explain and monitor the test, a perimeter and a specially designed instrument to check how wide the subject can see while he/she is focusing on a fixed point on a screen. This instrument displays a large number of stimuli varying in size and intensities and records the subject’s responses. Figure 1 illustrates a standard perimeter. In order to test the visual field, the clinician presents targets or stimuli to the patient in different parts of his visual field and ascertains whether he can see them or not. This is done in two basic methods: kinetic and static [4,5,7]. The kinetic strategy relies on the fact that the center of the visual field is normally more sensitive than its periphery. Such that a small and weak stimulus that is unviable at the edge of the visual field becomes visible as it is moved towards the fovea. For testing the visual field using this method, the perimetrist gradually moves a point of light from the periphery of the field towards the fovea and records the point at which the patient first sees the target. Kinetic perimetry is useful for mapping visual field sensitivity boundaries [5]. With the static technique, on the other hand, the clinician keeps the position of the stimulus constant and changes its intensity. It is known as the static technique because the target remains stationary. In this method each location in the visual field is tested at a number of different intensities to derive an estimate of its visual threshold. The recorded threshold, measured in decibel (dB), is the lowest intensity of the stimulus that can be reliably seen by the patient. Figure2 shows a sample of visual sensitivity thresholds for a normal eye. Static perimetry is done by automatic/computerized perimeters. Automated static perimetry is more sensitive to early visual field change than manual kinetic perimetry [3,4] ; also the data are quantifiable, reproducible and amenable to statistical manipulation. In addition, since it is computerized, the testing procedure is standardized and less dependent on who performs the test. Therefore for the most part it has been replaced kinetic perimetry There are several factors which affect the reliability of standard perimetry tests [4,5,7,8,9]: a) the patient’s inability to learn test strategy and maintain central fixation throughout the 30-45 minute test duration; b) fatigue and stress induced by the test c) patient’s false negative and false positive responses. Therefore, before interpreting visual field test results we should determine the accuracy of examination, or in other words, we must establish whether the results are reliable or not. One strategy is determining the test-reset reliability of individual thresholds [10]. Another way is based on the idea of predication/classification [11]. After the accuracy of the examination is assessed, the next step is to determine whether the field is normal or abnormal, and if it is abnormal what the disease is. Using Global indexes is one strategy to statistical analysis