Purchase this article with an account.
Antonio Ferreras, Blanca Monsalve, Pilar Calvo, Ana B. Pajarin, Laura Gil-Arribas, Michele Figus, Paolo Fogagnolo, Paolo Frezzotti, Luis E. Pablo; Discriminating Between Healthy And Glaucomatous Eyes With Spectral-domain Optical Coherence Tomography. Invest. Ophthalmol. Vis. Sci. 2012;53(14):660.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To design and validate a machine learning classifier for glaucoma diagnosis based on peripapillary retinal nerve fiber layer (RNFL) thicknesses measured with spectral-domain optical coherence tomography (OCT).
One hundred and eighty-eight healthy controls and 180 glaucoma patients were imaged with Cirrus optical coherence tomography (Carl Zeiss Meditec, Dublin, Ca). Glaucomatous eyes had intraocular pressure higher than 21 mmHg and abnormal standard automated perimetry (Humphrey Field Analyzer, Carl Zeiss Meditec; 24-2 SITA Standard strategy). Only one eye per participant was randomly included in the statistical analysis. Left eye data were converted to a right eye format. A linear discriminant function (LDF) was calculated according to the stepwise logistic regression results of 256 RNFL thickness values of the circle scan. The diagnostic accuracy of the LDF and other parameters included in the software of the Cirrus OCT were evaluated in another independent population.
Age and central corneal thickness did not differ significantly (p>0.05) between the groups in both samples. In the validating set, mean deviation of Humphrey perimetry was -6.89±5.9 dB in the glaucoma group. The greatest area under the receiver operating characteristic curves (AUCs) were 0.980 (95% confidence interval: 0.964-0.996) for the LDF, followed by the inferior quadrant thickness (0.926; 95% confidence interval: 0.885-0.967) and the average thickness (0.899; 95% confidence interval: 0.850-0.949). Significant differences were found between the AUC of the LDF and inferior quadrant thickness (p=0.001) and average thickness (p<0.001) using the DeLong method. The best sensitivity-balance was found for the LDF: 88%-99%, respectively.
In the teaching and validating sets, the LDF improved the diagnostic ability of OCT-provided parameters. Machine learning classifiers may increase diagnostic ability of OCT. These tools should be used in combination with the rest of OCT parameters and clinical examinations.
This PDF is available to Subscribers Only