MICROANEURYSMS DETECTION IN RETINA FUNDUS IMAGES WITH MULTIRESOLUTION GLOBALNET TO SUPPORT THE DIAGNOSIS OF DIABETIC RETINOPATHY

• Main Author: Ramadiastri, Fadhila
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/68835
• Institution: Institut Teknologi Bandung

Diabetic retinopathy is a complication of the eye from diabetes mellitus, which is one of the leading causes of blindness in the world. Due to the symptoms that do not appear in the early stages, the diagnosis for diabetic retinopathy patients are often late. Early diagnosis of diabetic retinopathy can be made by utilizing clinical indicators of retinal microaneurysms. Microaneurysms (MA) are pocket structures that appear in retinal blood vessels as a complication of diabetic retinopathy in the eye. In this study, the detection of MA with segmentation to support the diagnosis of diabetic retinopathy was carried out using a deep learning method based on the U-Net architecture called GlobalNet. This study performs semantic segmentation of microaneurysms using multi-resolution GlobalNet, which is used for various input image resolutions. Changes in the size of the input image is done using the bicubic interpolation method. In addition, this research also develops a method to combine information at the pixel level from outputs of the multi-resolution GlobalNet model. The evaluation was carried out at the pixel level for the GlobalNet single resolution segmentation model where the AUPR value was 0.391 ± 0.026 for the model with the input image resolution of 640 x 640 pixels (px), the AUPR value was 0.387 ± 0.035 for the model with the image resolution of 960 x 960 px, and AUPR value of 0.394 ± 0.050 for models with input image resolution of 1280 x 1280 px. The AUPR value of the research results still provides a lower AUPR value than similar research. The level evaluation provides sensitivity values for the GlobalNet segmentation model and the combination method with logic, where the average sensitivity values for models with input images of 640 x 640 px, 960 x 960 px, and 1280 x 1280 px are 0.689 ± 0.063, 0.659 ± 0.041 , and 0.647 ± 0.057 respectively, while the combination methods gives a sensitivity value of 0.672 ± 0.035. The sensitivity value at the lesion level has shown a higher sensitivity value compared to similar studies which also performed semantic segmentation of microaneurysms with U-Net based architecture.