Knowledge-Based Machine Learning for Glaucoma Diagnosis from Fundus Image Data

In recent years, the computer-aided diagnosis of glaucoma has seen great developmental strides. In this paper, we present algorithms for glaucoma diagnosis from fundus images by incorporating doctors' knowledge into algorithm development. We extract features of the fundus images from optic disc and optic cup boundary lines that were drawn by doctors, and from these features, predictions are made. To the optic disc and optic cup boundary lines, we meticulously divide the area based on doctors' knowledge and then conduct principal component analysis to extract the features. The extracted features correspond well with doctors' knowledge so that the diagnostic results can be intuitively explained, rather than just generate a black-box forecast. On a real sample set, the proposed feature extraction and diagnosis algorithms achieve good mean prediction accuracy.