Retinal Blood Vessel Segmentation using a Multi-Scale Layer in Deep Learning

Retinal Blood Vessel Segmentation in Fundus images is a field of study which relies on computational models to isolate the blood vessels from a Retinal Image. The primary difficulty addressed in past literature is the formulation of a uniform solution that can segment different types of blood vessels having different amount of contrast. Our method proposed in this paper is based on the approach of multiple patch size usage for segmentation using a Neural Network. The approach takes as input three different patch sizes for each pixel location which is passed through individual Convolutional Neural Network. The output from the three networks is translated into a 1-D feature vector using a novel Multi-Scale Layer in this work. The Multi-Scale Layer takes as input the feature vectors and combines the features on the basis of their importance in determining the classification of the central pixel of the patches into blood vessel or non-blood vessel pixel. The classification for each of the pixels in the image is finally combined to arrive at the final segmented image of the blood vessels in the Fundus image. The accuracy obtained for the proposed algorithm is equal to 94% in our work.