High Performance and Low Complexity Retinal Vessel Segmentation Method Based on Extended DCNN

The segmentation of the retinal vascular tree (RVT) is undoubtedly crucial for the detection and diagnosis of various pathologies that affect the retina. As a result, segmentation must have high performance with reduced complexity and execution time to meet constraints imposed by the clinical context. In this work, we propose a new method for RVT segmentation. The main contribution consists of suggesting a new DCNN, inspired from Segnet architecture where standard convolution blocks are substituted by MobileNet convolution blocks. The proposed method enhances segmentation performance and reduces computational complexity. The proposed method is tested on DRIVE database. It scored an Accuracy of 97.08 %, a sensitivity of 82.38 % and a specificity of 98.55%, within an architecture of 3.4 M parameters.