Sine-Net: A fully convolutional deep learning architecture for retinal blood vessel segmentation

Segmentation of blood vessels becomes an initial critical step in medical imaging because it is a key item for the diagnosis of many diseases in different fields including ophthalmology, neurosurgery, oncology, cardiology and laryngology. An automated tool for vessel segmentation can assist clinicians and con-tribute to patient treatment scheduling. However, it is still a challenging problem due to various condi-tions existing in images such as pathology, noise and poor contrast. Deep learning architectures typically achieve the-state-of-the-art performances in machine vision applications due to high contextual feature generations. This paper introduces a deep learning architecture for fully automated blood vessel segmen-tation. We propose a novel model, called Sine-Net, that first applies up-sampling and then down-sampling for catching thin and thick vessel features, respectively. We also include residuals to carry more contextual information to the deeper levels of the architecture. Deep networks may perform better if inputs are appropriately pre-processed. Thus, we conduct tests on our network with and without pre-processing applied to input images. We present experimental validations on retinal images of 3 publicly available databases (STARE, CHASE_DB1 and DRIVE) and compare results in terms of sensitivity, speci-ficity, accuracy and area under curve metrics. Our results demonstrate that Sine-Net outperforms the methods proposed in the literature in some of the metrics. In addition, the method has the potential to be used in clinical applications thanks to its decent execution time, high accuracy and robustness. (C) 2020 Karabuk University. Publishing services by Elsevier B.V.

Automated summary

This paper introduces a deep learning-based blood vessel segmentation method called Sine-Net, which captures different types of vessel features through upsampling and downsampling, and transmits more contextual information through residuals. Experimental results show that Sine-Net outperforms methods in the literature in some metrics and has the potential for clinical applications.