Abstract
Purpose :
Recent studies have shown that retinal oximetry could assist in the management of diseases such as diabetic retinopathy, central retinal vein occlusion, glaucoma and Alzheimer's disease.
In order for retinal oximetry to be automated, the retinal vessels must be segmented and classified into arteries and veins.
Here we demonstrate the use of convolutional neural networks on images from the OxymapT1 retinal oximeter to segment the retinal image into arteries and veins.
Methods :
Retinal oximetry images were obtained from 56 individuals who were healthy or suffered from a pathology of the eye with only one image from each individual being used. Each image was manually segmented into arteries and veins by an expert.
The neural network used for the segmentation task is based on the fully convolutional U-net architecture which has been successfully employed in many biomedical segmentation tasks.
Image preprocessing was applied order to correct for uneven illumination and in order to emphasize the difference between arteries and veins before the images were presented to the network for training. The performance is compared with results from the literature on the commonly used DRIVE dataset.
Results :
The proposed convolutional neural network was tested on the tasks of segmenting the retina into vessels and on how well it distinguished between arteries and veins.
The network achieves an accuracy score of 0.987 and an area under the ROC curve of 0.9899 when performing vessel segmentation and an accuracy score of 0.9743 and area under the ROC curve of 0.9919 when classifying vessels into arteries and veins.
Figure 1 compares the performance of the proposed method with results obtained conventional fundus images.
Figure 2 shows example output from the system.
Conclusions :
The system presented here performs at least as well in the task of vessel segmentation and separation of vessels into arteries and veins on images from the OxymapT1 as state-of-the-art systems do on the conventional RGB retinal images available through public datasets such as DRIVE.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.