MACHINE LEARNING ON RETINAL IMAGE FOR RETINAL DISEASE DETECTION

• Main Author: Sonya Tarabunga, Dewita
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/47231
• Institution: Institut Teknologi Bandung

There are various retinal diseases that are very dangerous and also can cause blindness if they are not detected early. To help eye specialists on retinal disease detection, we need an automated retinal disease detector model based on a retinal image. But currently, most of the researches focus on detecting specific retinal disease. This will cause problems if the patient does not know what disease they might currently have or any general patient doing some general eye check. Because of that, this final thesis is going to work on building a general retinal disease detection model. We use machine learning as a method to help solve the problem of automated detection from a retinal image. Detection model used here is using an initial weight gained from a transfer learning algorithm from other models that are already trained on other dataset such as ImageNet. We also use multiclass classification because we use three different retinal diseases and also to account for normal eyes without any disease. Experiments done on this thesis have shown a result that using deep learning will give a great performance on automated detection of a retinal disease from a retinal image. Moreover, Inception model gives the best result with the highest accuracy and a good recall value of every disease. The best accuracy gained from the experiment is 93%, gained by freezing some of the first layers of the model and a 0.0005 value of learning rate. The aforementioned model gives a recall for AMD, diabetic retinopathy, and glaucoma respectively 0.91, 0.94, and 1.