Retina Blood Vessel Segmentation Using Deep Learning

Abstract

The retinal image analysis is important in the early diagnosis of ocular conditions like diabetic retinopathy, glaucoma, and hypertension. The diagnostics of retinal blood vessels is tedious, and also requires professional skills to be identified by hand, which is why automated options are highly appreciated. An adapted U-Net design is suggested in the paper to introduce a deep learning-based retinal blood vessel segmentation approach. The model is trained with DRIVE dataset to supplement the quality of images using new preprocessing techniques like normalization and contrast enhancement. In order to enhance the performance of segmentation, hybrid loss that includes Dice Loss and Binary Cross Entropy with a combination of the two is used, which allows detecting both fat and skinny vessel structures. The proposed model has an accuracy of 95.0, Dice coefficient of 87.5, sensitivity of 84.2 and specificity of 97.1. The experimental results indicate that the model is better than the traditional methods and offers credible segmentation results. The paper shows that the deep learning techniques can be useful in automated retinal image analysis and may be applied in the clinical diagnosis.