Purpose
To apply computational methods in the discovery of 3D optic nerve head (ONH) structural phenotype features for detecting and monitoring glaucoma damage, and the discovery of new phenotype-genotype associations.
Methods
A subset of participants (n=370) from the Ocular Hypertension Treatment Study was selected on the basis of availability of simultaneous stereo fundus images. A stereo correspondence algorithm, optimized for fundus images, was applied to the set of stereo fundus pairs to produce a disparity map that quantitatively measured the ONH structure for each subject (Figure 1). Principal component analysis (PCA) was applied to the disparity maps to extract computational 3D ONH structural features. The first 25 principal components, or features, were retained and examined individually in building a predictive regression model for horizontal cup-to-disc ratio (HCDR). The relationship between the ONH features and demographic variables gender, age and ethnicity were also examined. In all cases, Bonferroni correction was used to adjust for multiple hypothesis testing.
Results
Five of the 25 computational 3D ONH features were significantly associated (corrected p<0.05) with HCDR, the association for a single ONH feature is shown in Figure 2. Combined, these features explained 65% of the variance in HCDR in the subjects. Significant associations were also found between the features and ethnicity, while suggestive associations were found with age and gender.
Conclusions
Using computational methods, we generated a set of structural features for quantifying the 3D shape of the ONH. These statistically independently features had significant association with and predictive power for HCDR, a clinically important measurement used to diagnose and monitor glaucoma. Intriguing associations of ONH phenotype features were also found with ethnicity, age and gender. Future work will explore the power of applying these features to detect and track glaucoma.
Keywords: 629 optic nerve •
549 image processing •
473 computational modeling