Learning Representations Through Ensemble Of Fuzzy C-Means For Identification Of Retinal Pathologies

Retinal image analysis is employed to automate screening process through low-level feature extraction and classification. Supervised classification approaches are dependent on kernels or distance metrics to handle complex manifolds as they warp feature space for effective classification with less complex boundaries between classes. Proposed approach identifies control points (Voronoi diagram) by exploring the structures of class specific manifolds which constructs complex boundaries with piecewise linear nature. Such a framework has less number of hyperparameters to tweak resulting easy control and understanding of the system. The learning characteristics of the proposed algorithm has been depicted on toy and optical coherence tomography data set. It has illustrated effective performance in identification of retinal pathologies and compared against off-the-shelf classifiers with various parameters. Proposed algorithm is capable of accommodating unsupervised approaches other than Fuzzy C-Means reflecting its adaptability.