Convolution Block Extension of DCNN for Retinal Vascular Segmentation: Taxonomy and Discussion

Abstract

The retinal vascular tree (RVT) segmentation is a main step for diagnosing several ocular diseases. Higher accurate segmentation remains crucial to ensure a reliable disease detection and hence clinical treatment. Numerous standard deep learning (DL) architectures have been employed to segment the RVT regardless of the image field However, due to the intricate morphologies of vascular trees comprising fine and complex structures, those DL architectures failed to achieve high accuracy in retinal vessel segmentation. Therefore, several promising solutions have been developed to overcome these limitations, where their main contributions rely on adapting the convolution processing of deep convolutional neural networks (DCNNs) blocks with respect to the retinal vessels characteristics. In this paper, we present a review of extended convolution blocks within DCNNs for RVT segmentation from fundus images. Our main contributions remain on (1) Identifying the different principles extension of convolution blocks; (2) Proposing a taxonomy of convolution block extension, and (3) Analyzing and discussing the strengths and weaknesses of each extension type with respect to segmentation quality and database characteristics. The presented study allows a valuable recommendation for future research in the field of RVT segmentation based on DCNN.