Hybrid Encoder-Decoder Model for Retinal Blood Vessels Segmentation

Segmented retinal blood vessels play a significant role in the clinical analysis of retinal vascular structure, which helps detect any eye diseases to prevent untimely impaired vision. With the rate of vision impairments globally, there is a need for a fast-automatic retinal blood vessel segmentation model to aid early detection of DR before its rapid progression to the high-risk stage of vision loss and blindness. The traditional UNET network has demonstrated steep success in biomedical segmentation tasks but is limited by the complexity of long training time due to many parameters. This paper proposes a hybrid encoder-decoder model based on the VGG16 encoder as the backbone and U-Net decoder with transfer learning for retina blood vessel segmentation to leverage the drawbacks. This approach aims to modify the traditional UNET architecture to optimize the training time and minimize computational cost and process complexities. The proposed framework resolves the limitation of long training and execution time compared with some U-Net based models, alleviates the complexity of high parameters, reduces computational resources and cost, minimizes loss, and alleviates overfitting. The evaluation of the proposed model on the DRIVE dataset obtains a promising result.