Automatic Image Segmentation by Ranking Based SVM in Convolutional Neural Network on Diabetic Fundus Image

Considering the diabetic retinopathy has been commanding as an emerging research field in ophthalmological disease diagnosis, the exact segmentation of the optic disc, fovea, and blood vessels have became essential level for automated diagnosis practices. In diabetic retinopathy, the fundus regions are normally overbright, faint regional boundary and irregular in shape. Besides, the features of fundus region vary from regular tissues and hence, training fundus images with an original convolutional neural network (CNN) is not enough. This research is motivated by state-of-the-art ranking support vector machine with CNN, utilization of a score function which is more suitable for multi-level classification to binary features classification to reduce the overall execution in the segmentation of fundus images. A two stage deep learning approach have been proposed in this chapter. In the first stage, a deep ranking support vector machine (SVM) is used to define a consistent feature label, which will be served as input to a CNN. It reduces the number of channels in the following CNN and allows it to converge on extended detailed segmentation of the optic discs and vessels. In order to gain the accuracy of the segmentation, the output constrained layer of SVM to CNN have been added so that the detector can derive the features of enclosing region in which the fundus regions are residing. It gives assess on widely used fundus datasets and show encouraging results over competing state-of-the-art methods.