Neural Network Analysis of Different Segmentation Strategies of Nerve Fiber Layer Assessment for Glaucoma Diagnosis.

PURPOSE: To compare the diagnostic performance of different segmentations of the nerve fiber layer (NFL) thickness measurements using an artificial neural network and to define the optimal number of sectors with best diagnostic ability for glaucoma diagnosis.
METHODS: A total of 117 glaucoma patients and 123 normal subjects were included in the study. NFL thickness measurements were performed using the Spectralis-OCT (Heidelberg Engineering) to obtain the NFL thickness average; measurements from 2 semicircles, 4 quadrants, and 6, 8, 12, 16, 24, 32, and 64 sectors; and 768 uniformly divided locations around the peripapillary NFL. An artificial neural network evaluation was performed to compare the influence of sector analysis on the diagnostic performance of optical coherence tomography. Receiver operating characteristic curves were used to compare the diagnostic ability of the different segmentation analyses.
RESULTS: The 6 sectors divided by the horizontal division of the nasal and temporal quadrants were better than the 6 sectors divided by the vertical line through the superior and inferior quadrants [areas under curve, 0.778; 95% confidence interval (CI), 0.720-0.829 and 0.814; 95% CI, 0.759-0.861, respectively]. In the case of quadrants, clock quadrants (area under curve 0.770; 95% CI, 0.712-0.822) were better than the ISNT (inferior-superior-nasal-temporal) quadrants (area under curve, 0.770; 95% CI, 0.712-0.822; P=0.003). The first segmentation strategy that improved the diagnostic value of 4 ISNT quadrants was the 12-sector analysis (area under curve, 0.845; 95% CI, 0.793-0.889; P=0.001).
CONCLUSIONS: The 2 best candidate strategies for the OCT report were the 12-sector analysis and the 4 planimetric quadrant (alternatively, the 4 clock quadrants) analysis.