Purchase this article with an account.
F. K. Horn, C. Y. Mardin, D. Baleanu, R. Laemmer, N. Bellios, A. M. Juenemann, R. P. Tornow, W. Adler; Analysis of Peripapillary Nerve Fiber Layer Thickness (Assessed by SOCT) With Automated Classification and Correction for Individual Optic Disk Locations. Invest. Ophthalmol. Vis. Sci. 2010;51(13):260.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To study the diagnostic value of retinal nerve fiber layer ring profiles measured peripapillary with spectral domain optical coherence tomography (SOCT) under consideration of the position of the optic disk in relation to fixation.
Thickness values of the retinal nerve fiber layer were determined in 768 locations around the optic disk using high resolution SOCT (Spectralis, Heidelberg Engineering). In all calculations the measured position of the optic disk in relation to the fixation was taken into account by an individual correction of the TSNIT-graph considering the fovea-disk-angle. Different sizes of sectors were used for training of two different state-of-the-art machine learning algorithms (random forest , support vector machine ): all 768 thickness measurements, quadrants, 32 sectors, and 6 nerve fiber bundle related areas. Classification performance was estimated by subject-based 10-fold cross-validation . Patients and subjects: 337 eyes from 232 patients with different stages of open angle glaucoma were included. 433 eyes from 269 healthy subjects served as control cohort. Controls and two patient groups (preperimetric and perimetric glaucomas) were compared using ROC-curves (one eye/patient).
Mean angle between the fovea and the center of the optic disc versus the horizon as measured by the present SOCT was 5.9±3.7°. ROC-curves for single peripapillary sectors indicate an increase of the validity when the position of the optic disk during SOCT-measurement was taken into account. Comparing controls and glaucoma patients, the best separation could be achieved if analysis was performed with a machine learning classifier using all data of the TSNIT-graph. Both automated algorithms of this study yielded similar performance. At a fixed specificity (97%) automated classification was able to uncover 91% of the present perimetric glaucoma patients, compared to 80% achieved by total mean thickness (comparison of areas under ROC-curves: p<0.05). In preperimetic glaucoma patients, however, none of the present methods performed better than the total mean thickness of the nerve fiber layer.
Consideration of the optic disk location during image acquisition as well as automatic classification is helpful in glaucoma detection by SOCT measurements.
 Breiman, 2001,  Vapnik, 2000,  Brenning, 2008.
Clinical Trial: :
www.clinicaltrials.gov NCT 00494923
This PDF is available to Subscribers Only