Optic Cup Segmentation using U-Net Architecture on Retinal Fundus Image
Pulung Hendro Prastyo, Amin Siddiq Sumi, Annis Nuraini
Abstract
Retinal fundus images are used by ophthalmologists to diagnose eye disease, such as glaucoma disease. The diagnosis of glaucoma is done by measuring changes in the cup-to-disc ratio. Segmenting the optic cup helps petrify ophthalmologists calculate the CDR of the retinal fundus image. This study proposed a deep learning approach using U-Net architecture to carry out segmentation task. This proposed method was evaluated on 650 color retinal fundus image. Then, U-Net was configured using 160 epochs, image input size = 128x128, Batch size = 32, optimizer = Adam, and loss function = Binary Cross Entropy. We employed the Dice Coefficient as the evaluator. Besides, the segmentation results were compared to the ground truth images. According to the experimental results, the performance of optic cup segmentation achieved 98.42% for the Dice coefficient and loss of 1,58%. These results implied that our proposed method succeeded in segmenting the optic cup on color retinal fundus images.