ABSTRACT
The human eye is a vital organ of vision, which gives us the sense of sight, and is of utmost importance in all day to day activities. However, there are many diseases that affect the human eye and can lead to blindness. The major causes of blindness include glaucoma, diabetic retinopathy and other corneal and retinal infections. However, most of these conditions can be prevented with an early diagnosis. There are several techniques of diagnosis in digital health which can be used for early detection of blindness by analyzing fundus images. These includes K-Nearest Neighbor algorithm (KNN), Support Vector Machine (SVM) and Artificial Neural Networks (ANN). Artificial Neural Networks has the advantage of accuracy in problem-solving, adaptive nature of learning, fault tolerance and real time operational mode quality over others. This project presents an Artificial Neural Network (ANN) based system to detect Glaucoma and Diabetic Retinopathy by automatically analyzing fundus eye images of suspected patients. The implementation of this project was done in four (4) modules. Data acquisition was done via publicly available online fundus database, and the database used includes Hossein Rabbani, SPIE, and ORIGA-light fundus database. The acquired fundus images were passed through Preprocessing stage. The preprocessing techniques used were grayscale, histogram equalization, and thresholding. Feature extraction was done using Gray Levels Co-occurrence Matrix (GLCM) and four textural features were extracted from each fundus image in form of positive integers. The four features are: Contrast, Correlation, Energy and Homogeneity. These four features served as the input into a Back Propagation Neural Networlc.. The Back Propagation Neural Network (BPNN) is mostly popular for its ability to minimize errors in classification. The Hidden Layer configuration used in the BPNN for the training were two hidden layers and three hidden neurons each (fed into each layer) at a threshold value •o.6'. The developed system was trained on a total of 135 fundus images, while 60 fundus images were used for proper testing and evaluation of the developed system. An accuracy of 90% was obtained for Diabetic Retinopathy; 95% obtained for Glaucoma; and 100% obtained for the Healthy class. The overall accuracy for the developed system was found to be 95%, coupled with the fact that it only takes about 13 seconds to screen one eye; as opposed to manual screening which takes about 15 to 20 minutes. The developed ANN-based eye diseases decision support system will assist in screening diabetic retinopathy and glaucoma. This will greatly assist ophthalmologists with early diagnosis; Hence, the system has the potential to reduce the possibility of vision loss or blindness.
