Classification of arteries and veins in montaged wild-field OCT angiograms using convolutional neural network

Abstract

Purpose : We propose to use a deep-learning-based method to differentiate arteries from veins in montaged wild-field optical coherence tomography (OCT) angiograms.

Methods : 6×6-mm disc, macular and temporal OCT angiography (OCTA) scans with a 400×400 A-line sampling density were acquired separately on 78 eyes (including 53 eyes with diabetic retinopathy (DR) and 25 healthy controls) using a 70-kHz commercial OCTA system (RTVue-XR; Optovue, Inc.). Angiogram at each region was generated by maximum projection of OCTA signal within the slab of inner retina, and then montaged into a wide-field angiogram covering 6x17-mm field of view. We proposed an end-to-end convolutional neural network that classifies arteries and veins (CAVnet) [Fig. 1] in montaged wild-field OCT angiograms. This method takes the OCTA images as input and outputs the segmentation results with arteries and veins identified. We not only classified arteries and veins down to the level of precapillary arterioles and postcapillary venules, but also detected the intersection of arteries (or arterioles) and veins (or venules). We evaluated the agreement (in pixel) between segmentation results and the ground truth using sensitivity, specificity, F1-score and Intersection over Union (IoU).

Results : We evaluated the average segmentation accuracy on test dataset [Fig. 2]. Our algorithm achieved an average sensitivity of 97.05%, specificity of 99.69%, F1 score of 96.20% and IoU of 92.79% on all arteries. Our algorithm achieved an average sensitivity of 95.71%, specificity of 99.80%, F1 score of 96.17% and IoU of 92.75% on all veins. We also achieved an accuracy of 79.94% identifying intersection points. The sensitivity of arteries is slightly higher than that of veins, which may because there is obviously capillary-free zone near the retinal arteries in OCTA images. Shown by all scans from test dataset, the performance (F1 score) of our algorithm was independent of scan quality, measured by signal strength index (Pearson correlation, p-value = 0.793). The results show CAVnet has high performance on differentiating arteries and veins even in severe DR cases.

Conclusions : CAVnet can classify artery and vein and their branches with high accuracy, and has potential application in the analysis of diseases such as branch retinal artery occlusion and branch retinal vein occlusion.

This is a 2021 Imaging in the Eye Conference abstract.