Abstract
Purpose :
To combine quantitative parameters of optical coherence tomography (OCT) and OCT angiography (OCTA) as a predictor for classification of normal (C1), type 2 diabetes mellitus without retinopathy (C2) and diabetic retinopathy (C3) using artificial intelligence (AI).
Methods :
OCTA scans of C1 (n=38,), C2 (n=49), and C3 (n=40 [13 - mild non-proliferative DR; 12 - moderate non-proliferative DR; 15 - proliferative DR] ) eyes with mean age: 50,54 and 56 respectively were acquired for a 3 mm × 3 mm scan area. Local fractal analysis was performed on superficial and deep (after projection artifact removal) images to estimate the vascular parameters (vessel density, spacing between large and small vessels). Sectoral analysis of parameters was performed in a central 1.5 mm diameter region excluding the foveal vascular zone (FAZ). Virtual averaging and retinal layer segmentation were performed using graph search. Then, OCT parameters were calculated from OCTA B-scans: inner and outer retinal volume, retinal volume and thicknesses. OCT and OCTA parameters were coupled using algorithms such as the random forest (RF), and support vector machine (SVM). Efficiency of these algorithms were tested using Area under the curve (AUC), classification accuracy (CA), F-1, precision (Pr), recall (Re), sensitivity (Se) and specificity (Sp).
Results :
Figure 1 show examples of vascular parameters (OCTA) and segmented layers (OCT). When OCT parameters alone were used, AUC of C1 (0.86) was significantly higher (p<0.05) in RF compared to SVM (0.79). However no significant (p>0.05) changes between RF and SVM were observed in C2 and C3 (Table 1a). When OCTA parameters of superficial and deep layer were analyzed separately, no significant (p>0.05) improvement was observed in AUC's compared to OCT parameters. Nevertheless, SVM (0.83) had higher (p<0.05) AUC than RF (0.75) for C1, when deep OCTA parameters were used (Table 1b). When OCT and OCTA parameters were combined, AUC's improved (p<0.05) for SVM and RF compared to using OCT and OCTA parameters alone. Overall for both Superficial and deep layer , SVM was the best classifier for C1, C2 and C3 as it had higher Se and Sp compared to RF (Table 1c).
Conclusions :
A novel approach to combine OCT and OCTA parameters of same scan area using AI improved the segregation of diabetic eyes with and without retinopathy. SVM was the best classifier of these eyes.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.