DETECTION OF STRUCTURAL DAMAGE FROM GLAUCOMA WITH CONFOCAL LASER IMAGE-ANALYSIS

Purpose. To determine which structural optic nerve head parameters mea sured with confocal scanning laser image analysis that best discrimina te between normal persons and those with glaucoma. Methods. One random ly selected eye of 53 patients with early open-angle glaucoma (average visual field mean deviation=-4.8 dB) and of 43 age-, race-, and refra ctive error-matched normal subjects were studied. The performance of n ine structural measures was evaluated with linear multivariate analysi s and a neural network: cup area, cup to disc area ratio, rim area, he ight variation contour, cup volume, rim volume, cup shape measure, mea n retinal nerve fiber layer thickness, and retinal nerve fiber layer c ross-section area. A discriminant function was derived with two thirds of the sample and its discriminant power tested on the remaining one third. This was repeated twice so that the entire sample was used for training and testing. A neural network was trained and tested in the s ame way. Stereoscopic color optic nerve photographs of the same eyes w ere evaluated qualitatively by three experienced, masked observers. Re ceiver operating characteristic (ROC) curves of discriminant function, neural network results, and qualitative evaluation were plotted. Comp arisons of the areas under the ROC curves were performed with nonparam etric statistics. Results. There were statistically significant differ ences between the normal and glaucoma groups for all measures (P less than or equal to 0.007) except for height variation contour, mean reti nal nerve fiber layer thickness, and retinal nerve fiber layer cross-s ection area, Cup shape measure provided the single best measure to dis tinguish between normal subjects and those with early glaucoma and had a diagnostic precision of 84%. Neural network diagnostic precision, w hen all measures were used, was 92% and decreased to 82% when cup shap e measure was omitted. The area under the ROC curve when all measures were combined was 0.94; it was significantly lower (P=0.04) when cup s hape measure was omitted (area=0.84). The area under the ROC curve for qualitative optic disc evaluation by experienced observers was 0.93. There was no statistically significant difference between qualitative evaluation and neural network performance (P=0.80). Conclusions. Cup s hape measure, the statistical third moment of the distribution of dept h values of the optic nerve head obtained with confocal laser image an alysis, can be used to discriminate between normal persons and those w ith early glaucomatous damage with high diagnostic precision.