A DEEP LEARNING APPROACH FOR MULTICLASS CLASSIFICATION OF RETINAL IMAGES

• Main Author: OON, WENG WAI
• Format: Final Year Project Report
• Language: English; English
• Published: Universiti Malaysia Sarawak, (UNIMAS) 2023
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://ir.unimas.my/id/eprint/44192/1/OON%20WENG%20WAI%20%20%2824%20pgs%29.pdf; http://ir.unimas.my/id/eprint/44192/5/WENG%20WAI%20%20ft.pdf; http://ir.unimas.my/id/eprint/44192/

The eyes fulfill a vital function in daily life as the primary sensory organs for perceiving the world. Neglecting proper attention and care for eye health can lead to vision loss or impairment, with severe eye diseases having the potential to cause complete blindness. The main problem addressed in this study is the global burden of eye diseases, worsened by a shortage of ophthalmologists and insufficient training in using essential instruments. These factors contribute to delays in diagnosis and treatment, underscoring the necessity for an automated approach to classify eye diseases from retinal images. Therefore, this study aims to compare various types of deep learning models, such as Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) to classify specific eye diseases, including cataracts, diabetic retinopathy, glaucoma, and normal classes. Based on a review of related research, it is evident that CNN is well�suited for image classification. Therefore, this study will train models using a proposed CNN, VGG19, and InceptionV3 architectures. In this study, the dataset collected from Kaggle underwent preprocessing involving histogram equalization and image segmentation techniques. It was then divided into training (80%), validation (10%), and testing (10%) sets. Performance evaluation of the models will be based on accuracy, precision, and recall. Techniques such as early stopping, model checkpointing, hyperparameter tuning, and fine-tuning were employed to optimize the models and improve their accuracy. VGG19 Model 3 with 16 convolutional layers, 3 dense layers (each with 300 neurons), a dropout rate of 0.5, and Adam optimizer with a learning rate of 0.0001, achieved impressive training accuracy of 90% and testing accuracy of 87%. The utilization of deep learning models in clinical settings could help improve the accuracy and speed of eye disease classification, leading to timely interventions and improved patient outcomes.