Characterizing OCTA Images with Gray Level Co-occurrence Matrix (GLCM) Statistics and Their Correlations to Capillary Density

Abstract

Purpose : Gray level co-occurrence matrix (GLCM) is a statistical analysis approach examining the spatial relations between pixels within an image. This approach provides information about the texture of an image. Using GLCM statistics, we characterized OCTA images in healthy subjects and patients with various stages of diabetic retinopathy and investigated their correlation with capillary density.

Methods : 30 controls and 75 diabetic eyes (30 no retinopathy - NoDR; 22 nonproliferative diabetic retinopathy - NPDR; and 23 proliferative diabetic retinopathy - PDR) were imaged using a commercial SDOCT system (Avanti RTVue-XR; Optovue). A 3x3 mm macular OCTA scan was obtained for each subject. Individual axial length was obtained for the correction of retinal magnification. Full vascular layer OCTA which included both large blood vessels and capillaries located between the inner limiting membrane and 70-µm below the posterior boundary of the inner plexiform layer, was used for image processing and analysis. Capillary images were created after the removal of large blood vessels on the full vascularlayer OCTA. Capillary images were then characterized using GLCM statistics using Matlab (The MathWorks Inc., Natick, MA). GLCM was computed by analyzing pairs of horizontally adjacent pixels with the OCTA image intensity scaled to 8 gray levels. Correlation between GLCM metrics (contrast, correlation, energy, and homogeneity) and capillary density (%) was performed using Pearson’s correlation analysis.

Results : As GLCM-contrast increased significantly (p<0.0001, r=0.87) with increasing capillary density, GLCM-energy (p<0.0001, r=0.90) and GLCM-homogeneity (p<0.0001, r=0.92) decreased significantly with increasing capillary density. There was no significant correlation observed between GLCM-correlation and capillary density (p=0.1, r=0.16).

Conclusions : GLCM statistics provide an alternative approach for direct and quantitative assessments of OCTA, which may reveal additional features of clinical images. Further exploration of their potential application for diagnostic imaging is warranted.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.