Arterioles and Venules Classification in Retinal Images Using Fully Convolutional Deep Neural Network

The abnormalities in size, shape and other morphological attributes of retinal vasculature have been prospectively associated as a physio-marker and predictor of many microvascular, systemic and ophthalmic diseases. The progression of retinopathy has a very different evolution in venules and arterioles with some biomarkers associated with only one type of vessel. The robust classification of retinal vasculature into arteriole/venule (AV) is the first step in the development of automated system for analyzing the vasculature biomarker association with disease prognosis. This paper presents an encoder-decoder based fully convolutional deep neural network for pixel level classification of retinal vasculature into arterioles and venules. The feature learning and inference will be done directly from the image without the requiring the segmented vasculature as a preliminary step. The complex patterns are automatically learned from the retinal image without requiring the handcrafted features. The methodology is trained and evaluated on a subset of the images collection obtained from a population-based study in the UK (EPIC Norfolk), producing 93.5% detection rate. This proposed technique will be optimized further and may replace the AV classification module in the QUARTZ software which is developed earlier by our research group.