HARD EXUDATE SEGMENTATION IN VARIOUS RESOLUTION OF RETINAL FUNDUS IMAGES TO SUPPORT THE DIAGNOSIS OF DIABETIC RETINOPATHY

• Main Author: Afirrah Vasmaulidzra, San
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/67884
• Institution: Institut Teknologi Bandung

Diabetic retinopathy is a complication of diabetes that can cause vision loss. High blood sugar in retinal vessels causes blockage and leakage of blood vessels fluids. To reduce the risk of vision loss, detection of this disease should be done as early as possible by checking the retina for any clinical signs of diabetic retinopathy. The screening of diabetic retinopathy should be done regularly and will need a large number of ophthalmologists. The detection of clinical signs using computer will help this process. One of the clinical signs that characterize diabetic retinopathy is hard exudate. Hard exudate happens because fluids from blood vessels of the eye leaks. The presence of hard exudates has been associated with higher risk of vision loss and requires treatment to disappear. Therefore, detection and segmentation of hard exudates can help in reducing the risk of vision loss, detecting diabetic retinopathy and monitor the progress of treatment to remove hard exudate. Currently, there are many deep learning methods used to segment hard exudates using U-Net. In this study, a semantic hard exudate segmentation model was trained on IDRiD dataset using U-Net architecture by tuning the parameters of activation function, loss function, and dropout. The size of the input image will affect computation, the larger the image, the heavier the computational burden. There has been research that changes the image size into 640 × 640, 960 × 640, and 1440 × 960 sizes to segment hard exudate. The 1440 × 960 size produced a higher F1-score than 960 × 640, and a higher AUPR than the 640 × 640 input size. Research was conducted to determine the effect of using input sizes of 640 × 640 and 960 × 960. F1-score from each input size do not vary much, 640 × 640 resulted in 0.6925, 960 × 960 resulted in 0.6934. The p-value between the segmentation results of 80 images of both sizes for the F1-score, accuracy, sensitivity, specificity, PPV, and NPV metrics each yielded a value greater than 0.05, indicating no significant difference between the sizes and the results of hard exudate segmentation. The main cause of false positives in the segmentation results is the resampling done to restore the image to its original size. Other false positives are caused by misclassification of reflective parts of larger blood vessels and similar retinal pathologies such as drusen and soft exudate.