The accurate vessel segmentation and classification of the retinal vessels to arterioles and venous (AV) is can be seen as the first steps in development of automated retinal image analysis system for disease prognosis. We present a segment-wise accuracy judgment technique that is used on an optimized encoder-decoder based fully convolutional deep neural network for classification of retinal vasculature into arterioles and venules. The technique has optimized loss function calculation based on the majority voting of the pixel colors within the segment. This enhanced the deep learning classification results of pixels within each vessel segment. the contextual .The deep learning network is trained and tested on the newly created AV classification database, which is prepared as a standard benchmark reference for AV classification. The feature learning and inference will be done directly from the image without requiring the segmented vasculature as a preliminary step, and the complex patterns are automatically learned from the retinal image without requiring the handcrafted features. The segment-wise similarity measure creates a cleverer deep learning model by adding the contextual segment-level judgment in addition to the pixel-wise loss calculation.
The methodology is trained on the subset of images taken from publicly available MESSIDOR dataset. The MESSIDOR images can be downloaded from here.