Abstract
Purpose :
Quantitative analysis of blood vessels is important for the management of eye disease for which vessel segmentation is often a crucial first step. While this can be done manually, it is a time consuming task and can be subjective and labour intensive. Traditional approaches often fail when the image contrast is insufficient due to patient factors. The purpose of this work is to develop effective automated deep learning vessel segmentation techniques for blood vessels on the corneal limbus and fovea in fluorescein angiographic images.
Methods :
We propose a vessel segmentation method using fully convolutional neural networks (CNNs) also known as ‘U-Net’, an extension of the traditional CNNs developed for classification tasks, due to the impressive speed and results that can be achieved. We train the network on manually annotated ground-truth data to give vessel prediction values for each pixel. After training, we test the network on two different problem datasets: corneal limbus vessels (14 images); and foveal vessels (15 images). All the images were acquired using the Heidelberg HRA2 (Heidelberg Engineering, Heidelberg, Germany).
Results :
Two segmentation examples are shown in the Figure. Mean accuracies of 0.936 and 0.925 and area under curves (AUROCs) of 0.939 and 0.915 were achieved for the corneal limbus and foveal datasets respectively, demonstrating the excellent performance of the method and robustness of the prediction.
Conclusions :
Reliable automation of this complex task can save considerable amounts of time and improve disease management and diagnostic potential. This paves the way for complete, fully automated systems to be realised for diagnosing conditions such as diabetic retinopathy and identifying occurrences and severity of corneal neovascularisation.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.