Abstract
Purpose: :
To investigate the sensitivity and specificity of machine learning classifiers and Spectral Domain OCT (SD-OCT) for the diagnosis of glaucoma.
Methods: :
Sixty six glaucoma patients (POAG) and 43 healthy individuals were recruited at the Glaucoma Service of the University of Campinas, Brazil. Inclusion criteria for glaucoma patients were: POAG, IOP > 21mmHg, best-corrected visual acuity of 20/40 or better, spherical equivalent within 5.0 D, ≥ 40 years of age, 2 consecutive abnormal and reliable visual fields. Inclusion criteria for healthy individuals were : IOP < 21 mmHg, best-corrected visual acuity of 20/40 or better, spherical equivalent within 5.0 D, ≥ 40 years of age, 2 consecutive and reliable normal visual fields, open angle on gonioscopy. All patients underwent a complete ophthalmologic examination, achromatic standard automated perimetry (SAP Humphrey 24-2 SITA, Carl Zeiss Meditec, Inc, Dublin, California) and RNFL imaging with SD-OCT (Cirrus HD-OCT; Carl Zeiss Meditec Inc., Dublin, California). SD-OCT images were obtained with undilated pupils by the same observer (VGV), who was masked for the diagnosis. ROC curves were obtained for all SD-OCT parameters. Subsequently, the following machine learning classifiers were tested: BAG (Bagging), NB (Naive-Bayes), MLP (Multiple Layer Program), and CTR (Decision Trees) algorithms. Areas under the ROC curves (aROCs) obtained for each parameter and each machine learning classifier were compared.
Results: :
The mean age was 48.5 ± 9.1 years for healthy individuals and 61.1 ± 10.3 years for glaucoma patients (p<0.001). There was no significant difference between the control and glaucoma groups regarding intraocular pressure (IOP) (13.6 ± 2.8 mmHg and 13.8 ± 2.7 mmHg, respectively) (p=0.42), but glaucoma patients were using a mean number of 2.2 ± 1.1 medications to lower IOP. MD values were -6.2 ± 6.1 dB for glaucoma patients and -1.5 ± 1.5 dB for healthy individuals (p<0.001). SD-OCT parameters with the greater aROCs were: average thickness (0.810 - CI 0.717-0.883), inferior quadrant (0.801- CI 0.708-0.876), 7 o'clock position (0.815 - CI 0.723-0,887) and 11 o'clock position (0.852 - CI 0.765-0.916). aROCs with machine learning classifiers varied from 0.585 (CTR) to 0.829 (BAG). aROC obtained with BAG was not significantly different from the aROC obtained with the best SD-OCT parameter (p= 0.61).
Conclusions: :
In this series, machine learning classifiers did not improve the sensitivity and specificity of SD-OCT for the diagnosis of glaucoma.
Keywords: imaging/image analysis: clinical • nerve fiber layer