Segmentation of Retinal Blood Vessels in Fundus Images Using Fully Convolutional Networks

Segmentation is mandatory to analyze retinal blood vessel indication for diagnosis of various ophthalmic and systemic disorders. This paper takes a look at this problem, specifically, the application of Fully Convolutional Networks (FCNs) in this whole process, with a primary focus on the U-Net architecture. In this work, the authors conduct the experiments on three datasets, namely DRIVE, STARE and CHASE_DB1, and prepare the images for modeling through normalization, resizing, and data augmentation. Due to the binary structure of the task, the loss function of the proposed method is binary cross-entropy along with the Dice coefficient, and the model is trained with the Adam optimizer. The improvement scores that was obtained shows that the proposed model has a high accuracy, precision, recall and F1 score on all the datasets used and has proven to be better than other existing methods. The study establishes the usefulness of FCNs in segmenting the tree structure of the retinal blood vessel structures in uniterrupted manner, and presented promising improvement for early diagnosis and treatment in the clinical application. it emphasizes on the necessity of utilizing far improved deep learning architectures and more holistic training mechanisms in striving to create more accurate diagnosis support systems in medical image analysis.