Retinal Blood Vessel Segmentation using Focal Tversky Loss Function and U-Net

Abstract

The most important and challenging step in biomedical image possessing is accurate segmentation of region of interest. Accurate retinal blood vessel segmentation is also most import step for design screening algorithm for fundus diseases like diabetic retinopathy glaucoma etc. The algorithm related to blood vessel segmentation reported in literature have low accuracy due to present of pathology in fundus or irregular illumination of light during image acquisition. Recently deep neural networks are being extensively used to increase the accuracy. In this paper, image preprocessing techniques used to improve the fundus image quality before the use of well-known U-Net for segmentation of blood vessel. This paper also uses Focal Tversky loss function for classification. Proposed algorithm has been verified on DRIVE dataset and achieve the segmentation sensitivity, specificity, accuracy and AUC as 80.4%, 98.38%, 96.42% and 98.08% respectively.