Similarity-weighted linear reconstruction of anterior chamber angles for glaucoma classification

We present a reconstruction-based method, called Similarity-Weighted Linear Reconstruction (SWLR), for glaucoma classification from OCT images containing the anterior chamber angle (ACA). SWLR identifies the glaucoma type via linear reconstruction of the ACA region from similar reference images, a classification approach that has recently been shown in certain computer vision applications to yield higher accuracy than classifiers, as it does not rely on feature set quality and it makes specific use of examples that have a similar appearance. The performance of a reconstruction-based approach, however, is greatly affected by how accurately the test image aligns with the references. To address this problem, we present a low-rank decomposition scheme for orientation correction that exploits the symmetry of anterior chamber cross-sections. Together with other techniques for translational alignment, this orientation correction leads to improved reconstruction-based glaucoma classification. Tests on a large-scale clinical dataset show the proposed SWLR classification algorithm to outperform the state-of-the-art methods.