Exudate Segmentation for Diabetic Retinopathy Using Modified FCN-8 and Dice Loss

The severity of diabetic retinopathy can lead to blindness if undiagnosed and untreated. The presence of hard exudates is one of the diabetic retinopathy symptoms. Therefore, automatic segmentation of hard exudates can provide an important diagnosis for diabetic retinopathy. Due to the relatively small dimensions of the exudates and the availability of the optic disc that has similar color, the exudate segmentation is a challenge in itself. In this study, we propose a modification of the fully convolutional network model (FCN-8) by combining FCN-8 and shortcuts to improve the performance of FCN-8. Each shortcut consists of a convolutional layer and batch normalization to reduce input degradation. Prior to processing the hard exudates using a modified FCN-8, the optic disc was removed from the retinal image by detecting the area using Faster R-CNN based on the Alexnet architecture. For training and testing, we applied the IDRiD dataset to evaluate the performance of our proposed architecture. Experiments show that our proposed architecture provides accuracy, sensitivity, specificity of 98.18 %, 81.7%, and 98.37 % respectively. Our proposed method gives higher sensitivity compared to Autoencoder, U-Net, FCN-32, FCN-16, and FCN-8.