Abstract

SO.16.10

Automatic evaluation of optic nerve head images

Paul-Rolf Preußner1, Jochen Wahl1, Sibylle Haas-Brähler2, Peter Morfeld2
1Universitäts-Augenklinik Mainz; 2EVONIK, Frankfurt/Main

Objective
To present the opportunities, the restrictions and the current development state of the automatic evaluation of optic nerve head (ONH) images.
Methods
The single steps of the method consist of the recording of digital colored fundus images, finding of an approximate ONH location, finding of the vessel truncus, stepwise definition of the ONH margin, first by an elliptical approximation, then refined by a local margin accuracy improvement, segmentation of the found ONH area into 3 regions (vessels, atrophic area, vital rim), determination of sizes and centers of gravity of the 3 regions found and of the vertical/horizontal ratio of the atrophic region. For testing purposes, the procedure was applied to 8638 fundus images of 4319 employees of a German company (EVONIK).
Results
Depending on some threshold settings, approx. 1-3% of the images cannot be evaluated, as found by the software itself. In addition, approx. 0.5% of the evaluations are seriously wrong, without recognition of these failures by the software. Main sources of error are poor image quality (illumination, blurring, vignetting), but also “special features” of the individual fundus, e.g. unusual pigmentation. For the big majority of the images at least “plausible” results are obtained for which data correlations are in accordance with clinical experience, e.g. correlations between relative or absolute size of the atrophic region on the one hand and shift of the vessel truncus on the other hand, or of the data of fellow eyes with each other.
Conclusions
Currently, the method appears suitable particularly for the statistical evaluation of large data sets. Wether and at what time the boundary to an “average” or even an “experienced” investigator can be crossed remains to be seen in future.

Copyright: 2009 Porstmann Kongresse GmbH. All rights reserved. Impressum, Disclaimer.