Publication

Abstract : Retinal fundus images provide valuable diagnostic and clinical information in the diagnosis of ophthalmologic diseases. Retinal blood vessel analysis provides important diagnostic information about thinning of the retinal nerve fiber layer and alteration in the structural appearance of the optic nerve head. Here, an accurate retinal vessel detection method is proposed from fundus images using a generative adversarial network (GAN) utilizing multiple loss functions. The proposed GAN architecture consists of the generator as a segmentation network and the discriminator as a classification network. The generator is a multi-scale residual convolutional neural network with skip connection and up-sampling, while the discriminator is a vision transformer that acts as a binary classifier. The inception module extracts multi-scale features of vessel segments from different scales and captures fine vessel segments. The discriminator consists of stacked self-attention networks and position-wise fully connected feed-forward networks inferring two-class output. The attention mechanism in the transformer is competent to preserve both global and local information while acting as a discriminator. The proposed GAN model segments the blood vessels more accurately through the adversarial learning process to produce state-of-the-art results. In the preprocessing stage, the contrast of blood vessels is enhanced by contrast-limited adaptive histogram equalization algorithm. The robustness and efficacy of the proposed method have been evaluated on publicly available DRIVE, STARE, CHASE_DB1, HRF, ARIA, IOSTAR, and RC-SLO databases. Different performance measures like accuracy, sensitivity, precision, intersection of union, and F1Score are adopted to compare the proposed method with the existing methods available in the literature. The proposed method attains an accuracy of 0.9873 for CHASE_DB1 database, 0.9742 for DRIVE database, 0.9773 for HRF database, and 0.9628 for ARIA database.