Current keratoconus detection methods compared with a neural network approach.

Abstract

PURPOSE: Four videokeratographic methods for keratoconus detection were compared with a neural network approach. METHODS: A classification neural network for keratoconus screening was designed to detect the presence of keratoconus (KC) or keratoconus suspects (KCS); a separate cone severity network graded the severity of conelike topography patterns consistent with KC or KCS. Three hundred TMS-1 examinations (Tomey) were randomly divided into training and test sets. Ten topographic indexes were network inputs. Nine categories were used: normal, astigmatism, KC, KCS, contact lens-induced warpage, pellucid marginal degeneration, photorefractive keratectomy, radial keratotomy, and penetrating keratoplasty. KC was subdivided into KC1 (mild), KC2 (moderate), and KC3 (advanced). There were three outputs for the classification network (KC, KCS, and OTHER); target output values of 0 = OTHER, 0.25 = KCS, 0.5 = KC1, 0.75 = KC2, and 1.0 = KC3 were used for the severity network. RESULTS: The best-trained classification network had 100% accuracy, specificity, and sensitivity for the test set. The severity network had mean outputs (+/-standard deviation) of OTHER = 0.02+/-0.02, KCS = 0.21+/-0.05, KC1 = 0.52+/-0.17, KC2 = 0.74+/-0.12, and KC3 = 0.91+/-0.15. The severity network output for all categories was well correlated to the keratoconus prediction index (R = 0.892, P < 0.0001). The classification network had an overall accuracy and specificity significantly better (P < or = 0.005) than the Klyce/Maeda keratoconus index (KCI) test, the Rabinowitz test (K & I-S), and simulated keratometry (average Sim K). However, there were no significant differences in keratoconus sensitivity between the classification network, KCI, and K & I-S. The sensitivity and specificity of average Sim K were significantly worse than those of the other tests. The classification network had significantly better sensitivity (P < 0.001) and specificity (P = 0.025) for KCS detection than the K & I-S. CONCLUSIONS: The neural networks completely distinguished KC from KCS and from topographies that resembled KC. The network approach equaled the sensitivity of currently used tests for keratoconus detection and outperformed them in terms of accuracy and specificity.