Region Separated Vessel Segmentation in Fundus Image Using Multi-scale Layer-Based Convolutional Neural Network

Abstract

A region-based Multi-scale Convolutional Neural Network is used in this work to automatically segment the blood vessel pixels in Fundus images. Firstly, a region-based image partitioning method is implemented which separates each image into a set of three homogenous regions, namely Optic Disc, High Contrast and Low Contrast regions. kNN-based clustering approach is used for the image partitioning. Three Multi-scale Layer-based CNN models are then trained individually for each of the regions and tested on the DRIVE and STARE datasets. The obtained output is evaluated based on Accuracy, Sensitivity and Specificity, respectively, and the achieved results are reported in this work.