Glaucoma classification with a fusion of segmentation and image-based features

Automated classification of glaucoma is of interest in early detection and treatment. Existing methods employ features which are either image-based or derived from Optic Disc (OD) and Cup (OC) segmentation. While the latter suffers from segmentation inaccuracies, the image-based features tend to overfit in limited availability of training data. We propose a solution to overcome these issues and present a classification framework that fuses both type of features within a co-training based semi-supervised setting to overcome the paucity of labelled data. A novel set of features is proposed to represent the segmented OD-OC regions. Additionally, features based on Texture of projections and color Bag of Visual Words have been designed to be sensitive to the sector-wise deformations in OD. The proposed method was trained on 386 labelled and 717 unlabelled images. It outperformed existing methods with an accuracy and AUC of 73.28%, 0.79 on a private test set of 696 unseen images and 76.77%, 0.78 when cross-tested on DRISHTI-GS1 dataset.