Fovea Segmentation Using Semi-Supervised Learning

Despite the accessibility of retinal fundus images in recent years, fovea segmentation is an exacting task due to the insufficiency of labelled data. In this paper, we propose a deep learning pipeline which utilizes unlabelled data alongside labelled data for the segmentation of fovea. We train 484 labelled images using the Deeplabv3+ architecture and deploy EfficientNet-B3 as the encoder in the framework. Additionally, we introduce semi-supervised learning in our pipeline and train 1200 unlabelled images by generating their pseudo labels. We evaluate our results on the Jaccard index, Dice score, sensitivity, specificity and accuracy. Our Dice score of 82.43% and Jaccard index of 70.52% surpasses the existing methods. We obtain 91.74% sensitivity, 99.75% specificity and 99.57% accuracy.