Multiple Retinal Disease Classification system from Reti nal Images Usi

Abstract

Retina is a metabolically active part of the eye, in which eye specific and systemic based diseases manifest themselves in it. Retinal imaging with fundus camera is pri marily technique to diagnosis common causes of blindness and health problems of the eye, for its safety and cost effectiveness. Automated retinal disease classification to types and sub-types from fundus camera is essential for early detection and manage ment of disease treatment plan, which is less effective and tiresome with the manual methods. Different machine and deep learning methods have been proposed to au tomate such manual diagnosis procedures. However, the results reported from pre viously proposed automation techniques are inconsistent, less accurate, and did not consider Ethiopian context data. This thesis proposed the use of deep learning based classification of the four main causes of blindness (diabetic retinopathy, hypertensive retinopathy, glaucoma, and occlusion) to their respective types, as well as the subtypes of diabetic retinopathy and occlusion, using retinal fundus images. The model was developed using the Keras framework, trained, and validated using fundus photogra phy images acquired from the ”paper pointed dataset”, ”the Kaggle online repository” and local data acquired from Blue Vision Clinic. Google Collaborator and Python programming languages were used for model training and graphical user interface development. The ResNet50 based transfer learning model was used, and the result was then compared to benchmark finding. ResNet50 based transfer learning model showed a better result in type and subtype classification. With the proposed method, 97.2%, 96.8%, and 95.3% accuracy and of 93.2%, 92.2% and 93% F1-scores was achieved for disease type, diabetic retinopathy subtypes and occlusion sub-types respectively. These findings highlight deep learning based retinal disease classification as an effi cient and accurate alternative to manual methods, contributing to improved disease management and treatment.