Quantification of diabetic retinopathy using neural networks and sensitivity analysis

The design of neural network classifiers for the identification of diabetic retinopathy is discussed. Red-free digitised fundal images are tiled, and a neural network is trained to distinguish exudates from drusen (similar appearing lesions). By quantifying the degree of retinopathy, the approach can be used to screen diabetic patients for referral. A novel form of hierarchical feature selection using sensitivity analysis is presented. The resulting neural network is compact, and achieves 91% sensitivity and specificity on a test set.