Retinal vessel segmentation aims to separate the structure of blood vessels from the fundus image background. This retinal vasculature is then used for detection of numerous diseases like retinopathy of prematurity (ROP), glaucoma, diabetic retinopathy, coronary heart disease, etc. Deep learning-based semantic segmentation techniques are considered a breakthrough in the field of medical diagnosis using artificial intelligence. Various methods for segmentation with convolutional neural networks have been developed which have become indispensable in tackling more advanced challenges with image segmentation. The limitation is that they all require huge quantities of labeled data which is difficult to collect. To overcome this, U-Net architecture is widely used for segmentation of medical images as it segments the pixels individually and can be trained with a small number of images. In this work, we have implemented U-Net architecture and evaluated it on two public datasets: ‘HRF’ and ‘DRIVE’. An accuracy of 96.64% was obtained on the DRIVE dataset and 94.28% on HRF dataset. To check the model robustness, we tested the model trained on the augmented DRIVE dataset on the HRF dataset and vice versa. The model trained on the augmented HRF training set achieves an accuracy of 95.04% when tested on the DRIVE dataset. Similarly, the model trained on the augmented DRIVE training set achieves an accuracy of 92.17% when tested on the HRF dataset. A progressive web application is also developed as part of this work, since there is no specific easy-to-use interface to perform segmentation of retinal fundus mentioned in the literature, to the best of our knowledge. This application accepts retinal fundus image as input and performs segmentation using trained U-Net model, to provide an output image of the blood vessels which will assist the ophthalmologists in screening retinopathy.

Keywords

U-Net Deep learning Semantic segmentation Retinal image segmentation Blood vessel patterns Machine learning