Learning first-order rules from image applied to glaucoma diagnosis

Abstract

Computer-based diagnosis from image data is important for medicine. In particular, for the glaucoma diagnosis we target here, the ocular fundus image can be easily obtained and can be used to automatically identify whether an eye is glaucomatous or not. However, the image has a two-dimensional distribution, and it is difficult to feature the whole image through some real-valued parameters in general. This paper proposes a machine learning method using a set of expert’s decision cases that identify local abnormalities of an image. This method finds regularities between an image set and the decision cases using Inductive Logic Programming (ILP). Unlike decision-tree learning and neural networks, ILP allows relational learning between concepts. Learned rules are abstract enough to absorb noisy data obtained directly from image analysis. We applied the method to detecting early glaucomatous eyes. Our ILP system, GKS produced 30 rules from 2000 positive and negative examples that were obtained by segmenting 39 glaucomatous eyes. A 10-fold cross validation assessment shows about 80% sensitivity and 65% accuracy of the rules, resulting in the high performance comparable with human-level classification.