Visual field analysis using artificial neural networks

There have been several reports on the application of artificial neural networks (ANNs) to visual field classification. While these have demonstrated that neural networks can be used with good results they have not explored the effects that the training set can have upon network performance nor emphasized the unique value of ANNs in visual field analysis. This paper considers the problem of differentiating normal and glaucomatous visual fields and explores different training set characteristics using field data collected from a Henson CFS2000 perimeter. Training set properties including size, balance between normals and glaucomas, extent of field loss and the spatial location of glaucomatous defects are explored. A multilayer network with 132 input nodes, 20 hidden layer nodes and 2 output nodes is trained using an error backpropagation algorithm. Both sensitivity and specificity are measured during testing. The results demonstrate that large random sets are better than small random sets since sensitivity improves with size and specificity is not adversely affected. The variability in performance also reduces as training set size increases. In addition, sets that are biased towards glaucoma examples are more sensitive and less specific, while sets biased with normal examples are more specific and less sensitive than balanced sets. Thus large training sets with class balance are generally desirable for good sensitivities and specificities. The actual glaucoma examples contained in the set are also important, A training set deficient in examples of early loss produce poor sensitivities and good specificities, while an absence of advanced loss training examples has no detrimental effect on sensitivity or specificity. The spatial distribution of detects is also crucial. Spatially biased sets are unable to recognize defects that occur in locations where no previous defect has been presented while more balanced sets lead to better performance. In conclusion the ‘ideal’ training set should contain many examples of early defects that represent the full range of locations where these defects may occur.