Neural networks for visual field analysis: How do they compare with other algorithms?

Purpose: To compare the performance of a neural network in identifying visu al field defects with the performance of other available algorithms. Methods: A feed-forward neural network with a single hidden layer was train ed to recognize visual field defects previously collected in a longitudinal follow-up glaucoma study, and then tested on fields taken from the same st udy but not used in the training. The receiver operating characteristics of the network then were compared with the previously determined performance of other algorithms on the same data set. Results: At a specificity greater than 90%, the neural network was more sen sitive than any of the available algorithms (although only the global indic es were available for comparison, as the cluster and cross-meridional algor ithms did not achieve such high specificity at their current settings). At a lower specificity (80-85%), the neural network was unable to attain the h igh sensitivity of the cluster or cross-meridional algorithms; in fact, the cluster algorithm from the Low-Tension Glaucoma study was significantly mo re sensitive. Conclusion: The receiver operating characteristics of a feed-forward neural network designed to detect visual field defects were explored. At a very h igh specificity (90-95%) a neural network performed better than the global indices. However, at a lower specificity (78%-88%), the neural network perf ormed worse than cluster and cross-meridional algorithms.