Development of a new computational model supported by artificial intelligence for detection of retinal diseases

Retinal diseases are among the major health problems that cause serious vision loss, with diabetes and age-related health problems increasing worldwide. Early diagnosis and effective treatment play a critical role in preventing permanent vision loss caused by these diseases. Optical Coherence Tomography (OCT) imaging technology is widely used in diagnosis and treatment processes by non-invasively revealing microscopic structural changes in the retinal layer with high resolution. However, manual evaluation of these images is a time-consuming and specialized process. In this study, a new computational model supported by artificial intelligence is developed for automatic diagnosis of retinal diseases using OCT images. The system is built on a multi-layered and modular architecture that includes image validation, preprocessing, deep learning-based classification and decision support modules. The OCT-2017 open access dataset was used for model training, and four different retinal disease classes were targeted: CNV (Choroidal Neovascularization), DME (Diabetic Macular Edema), Drusen and Normal. Four different architectures were tested as deep learning-based classifiers (GoogleNet, ResNet50, EfficientNet-B0 and DenseNet-201). As a result of the comparative experiments, the DenseNet-201 architecture, which stands out with its dense inter-layer connection mechanism, achieved the highest accuracy rate and was preferred in the final version of the model. Training and test results were evaluated in detail with statistical metrics such as accuracy, sensitivity, specificity, F1 score and ROC curves. In addition, it is aimed to contribute to clinical decision support processes by presenting graphical and numerical prediction results to healthcare professionals through the user-friendly interface integrated into the system. Considering the integration potential of the developed system with health software and OCT devices, an expandable infrastructure has been created for future clinical applications. This thesis not only demonstrates the effectiveness of artificial intelligence-based decision support systems in the early diagnosis of retinal diseases, but also contributes to the literature with its unique data management, security and user interface components.