Abstract
Purpose :
The purpose of this study is to establish differential arterial-venous (AV) flow index analysis in optical coherence tomography angiography (OCTA) and validate it for early detection of diabetic retinopathy (DR).
Methods :
A convolutional neural network (CNN) has been developed to achieve automated construction of the OCTA-AV map. Based on morphological characteristics in OCTA (Fig. 1A), ground truth AV map maps were prepared (Fig. 1B). For generating AVA maps, the k-nearest neighbor (kNN) classifier was used to classify pixels as AV areas (Fig. 1C). By multiplying the OCTA image with the AVA map, the OCTA-AV map was constructed with flow intensity information preserved (Fig. 1D). The CNN was trained to achieve automated OCTA-AV construction (Fig. 1E and 1F). A 5-fold cross-validation was implemented. Quantitative flow index features, including arterial area (AA), venous area (VA), AVA ratio (AVAR), total perfusion intensity density (T-PID), arterial PID (A-PID), venous PID (V-PID), and arterial-venous PID ratio (AV-PIDR), were developed and validated for early detection of DR.
Results :
Figure 2 shows representative OCTA, ground truth AVA, predicted AVA, ground truth OCTA-AV, and predicted OCTA-AV maps. The CNN achieved an average IoU of 75.58%, a Dice score of 86.09%, and an accuracy of 84.68%. Quantitative analysis revealed that the area features AA, VA and AVAR can reveal significant differences between the control and diabetic eyes (NoDR and mild DR) but cannot separate NoDR and mild DR from each other. Vascular perfusion parameters T-PID and V-PID can differentiate mild DR from control and NoDR groups but cannot separate control and NoDR from each other. In contrast, the AV-PIDR can disclose significant differences among all three groups, i.e., control, NoDR, and mild DR.
Conclusions :
In this study, a deep learning network developed for robust AVA segmentation in OCTA images. The area features AA, VA and AVAR can reveal significant differences between the control and diabetic eyes. The PID features T-PID and V-PID can differentiate mild DR from control and NoDR groups. The AV-PIDR can disclose significant differences among all three groups, i.e., control, NoDR, and mild DR.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.