Retina blood vessels segmentation by combining deep learning networks

Abstract

In this paper, we propose two deep learning architectures for the segmentation and detection of the vascular networks of blood vessels in fundus images. First, we combined VGG16 with U-net, then, we used Resnet 34 in combination with U-net. Both architectures employ an encoding and a decoding path. In this paper, we used the DRIVE and STARE databases. After applying VGG 16+U-net on the DRIVE database, we obtained the accuracy value of 0.96955, 0.79929 sensitivity, 0.98624 specificity, 0.9805 recall, and 0.9833 F1-score. We applied VGG 16+U-net on STARE database and we got 0.95259 accuracy, 0.89996 sensitivity, 0.95530 specificity, 0.9933 recall, and 0.9742 F1-score. Concerning Resnet 34 + U-net, we got the value of 0.9692 accuracy, 0.7859 sensitivity, 0.9870 specificity, 0.9794 recall, and 0.9832 F1-score after applying on DRIVE database. Moreover, we got 0.9363 accuracy, 0.9335 sensitivity, 0.9246 specificity, 0.9961 recall, and 0.9649 F1-score after we applied Resnet 34+U-net on STARE.