Encoder Selection and Segmentation Quality in U-Net-based Retinal Vessel Analysis
Abstract
Retinal vessel segmentation through automation is a very significant area in retinal image processing, which can help in diagnosing many different ocular and systemic diseases. In this research study, a comparative study of the encoder-based U-Net models is presented for retinal vessel segmentation. Five different models including ResNet18+U-Net, ResNet34+U-Net, Traditional U-Net, VGG16+U-Net, and VGG19+U-Net have been developed and trained on the retinal images belonging to the DRIVE and CHASE_DB1 databases. The proposed methodology involves image pre-processing, image normalization, image augmentation, feature extraction from image using the encoder architecture, and U-Net decoding with the skip connection for pixel-wise segmentation of retinal vessels. Among all models evaluated in this paper, the VGG19+U-Net model gives the highest value of Jaccard score of 81.83%, F1 score of 89.53%, Precision of 89.47%, Accuracy of 98.41%, and Specificity of 99.13%, whereas the highest Sensitivity score of 90.32% is achieved by the VGG16+U-Net model.