An evolutionary U-shaped network for Retinal Vessel Segmentation using Binary Teaching–Learning-Based Optimization

Abstract

In medical imaging, Retinal Vessel Segmentation (RVS) plays a significant role in finding the pathological changes in retinal blood vessels that can be used to detect various diseases like arteriolosclerosis, high blood pressure, diabetes, etc. Recently, convolutional neural networks (CNNs) and U-shaped (encoder–decoder) based models have been widely used in RVS tasks. However, these segmentation models are developed manually, which is tedious, requires high expertise, error-prone, and lost in preserving micro-vasculature details at the end of vessels. In this paper, we developed a Binary Teaching–Learning-Based Optimization (BTLBO) based evolutionary model to discover the optimal block structures in the U-shaped network for RVS. The proposed model also optimizes the network structure dynamically using flexible search space. Furthermore, we adopted an attention mechanism to find the complex structure in the retinal image. Moreover, the implemented cache system can speed up the evolutionary process. Finally, we evaluated the proposed model named BTU-Net on five retinal vessel image datasets to show the model’s potentiality in discovering high-performance optimal network structures for the RVS task.