Abstract
Purpose :
As the rates for development of diabetic retinopathy (DR) continue to rise, innovative computational methods have been developed to improve screening. The purpose of the study was to report an ordinary least squares (OLS) modeling method applied to optical coherence tomography angiography (OCTA) images for classifying proliferative DR (PDR) and non-diabetic control (NC) subjects.
Methods :
Twenty-two NC and 13 PDR subjects participated in the study. OCTA images of the superficial capillary plexus were obtained in a 6 x 6 mm macular region centered on the fovea. Each image was considered a solution to a linear autoregressive equation and 25 coefficients were estimated from the OLS model. Vessel metrics, namely vessel tortuosity index (VTI), vessel density (VD), spacing between small vessels (SSV), and between large vessels (SLV) were quantified from OCTA images using established methods. Vessel metrics were compared between NC and PDR groups and correlated with OLS coefficients. Logistic regressions were performed to identify individual and multiple coefficients associated with groups. Performance, predictive ability, and validity of the multivariate logistic regression were evaluated.
Results :
Age and sex did not differ between groups (P ≥ 0.09). VD and SSV were decreased (P < 0.0001) and SLV was increased (P < 0.0001) in PDR compared to NC group. VTI did not differ between groups (P = 0.34). VTI and VD were each correlated with 4 coefficients (P ≤ 0.04; N=35). SSV and SLV were correlated with 3 and 5 coefficients, respectively (P ≤ 0.04; N=35). Six coefficients that were individually associated with groups showed improved performance in multivariate logistic regression (Likelihood Ratio X2 = 38.17, P < 0.0001) and correctly classified 94% of subjects. The area under the receiver operating characteristic curve (AUROC) was 0.99 and AUROCs with and without cross validation did not differ significantly (P = 0.07), thereby validating the model.
Conclusions :
Using OCTA images, an OLS modeling method generated coefficients that correlated with vessel metrics and were utilized to derive predicted probabilities of PDR and NC. With further development, the method may have potential clinical utility and contribute to image-based computer-aided screening and classification of stages of DR and other ocular and systemic diseases.
This is a 2020 ARVO Annual Meeting abstract.