INTERPRETATION OF AUTOMATED PERIMETRY FOR GLAUCOMA BY NEURAL-NETWORK

Purpose. Neural networks were trained to interpret the visual fields f rom an automated perimeter. The authors evaluated the reliability of t he trained neural networks to discriminate between normal eyes and eye s with glaucoma. Methods. Inclusion criteria for glaucomatous and norm al eyes were the intraocular pressure and the appearance of the optic nerve; previous visual fields were not used. The authors compared the backpropagation learning method used by automated neural networks to t hose used by two specialists in glaucoma to classify the central 24 de grees automated perimetric visual fields from 60 normal and 60 glaucom atous eyes. Results. The glaucoma experts and a trained two-layered ne twork were each correct at approximately 67%. The average sensitivity of this test was 59% for the two glaucoma specialists and 65% for the two-layered network. The corresponding specificities were 74% and 71% for the specialists and the two-layered network, respectively. The exp erts and the network were in agreement about 74% of the time, which in dicated no significant disagreement between the methods of testing. Fe ature analysis with a one-layered network determined the most importan t visual field positions. Conclusions. The authors conclude that a neu ral network can be taught to be as proficient as a trained reader in i nterpreting visual fields for glaucoma.