DETECTION AND CLASSIFICATION OF DIABETIC RETINOPATHY WITH DEEP LEARNING

• Main Author: Wijaya, Erick
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39830
• Institution: Institut Teknologi Bandung

Increasing cases of diabetes cause an increase chances of diabetic retinopathy. Examination of diabetic retinopathy by expert is still done manually and requires time and effort. More diabetic retinopathy will cause doctors to take more time to make a diagnosis. This final project aims to produce a classification and detection model that can be used to determine the severity of diabetic retinopathy. This will help doctors to speed up the diagnosis process so that they can focus on more important stuff such as prescribing patients. The system consists of three main components, namely preprocessing component, object detection component, and classification component. In the preprocessing component, the collection of images is processed using a method of reducing the local average color to reduce image variation and make the object of diabetic retinopathy appear clearer. The object detection component utilizes the Mask R-CNN approach to detect the location of signs of diabetic retinopathy. The classification component uses the CNN topology to classify the severity of diabetic retinopathy. The scenario carried out in the experiment to determine the best hyperparameter is a one factor at a time. The experiment results show that the object detection model built with the Mask R-CNN with ResNet101 topology and utilized hyperparameter adjustment can produce an average mAP of 0.42. This value is higher than the related research model with average mAP of 0.17. The classification model was built with VGG16 topology and produced an accuracy of 0.75, slightly better than the related study using InceptionV3 with an accuracy of 0.73.