Abstract
Eye movements, optical opacities, and other factors can introduce artifacts during the acquisition of optical coherence tomography angiography volumes, resulting in suboptimal imaging quality. We aim to develop an automated deep learning model to separate excellent-quality from suboptimal-quality volumes in a quantitative and objective manner. Existing works use supervised classifiers trained on 2D en face images, which 1) represent quality as rigid and discrete classes, 2) require large amounts of labeled data for every type of artifact to generalize effectively, and 3) discard valuable depth information from the original volume. We propose OCTA-GAN, an efficient 3D generative adversarial network architecture that incorporates multi-scale processing layers to assess the quality of scans by fusing fine vasculature details with larger anatomical context. The unsupervised model learns patterns associated with excellent-quality volumes and accurately determines the quality of unseen volumes. Experimental results show OCTA-GAN's discriminator distinguishes excellent-quality from suboptimal-quality volumes with an AUC of 0.92, a sensitivity of 95.7%, and a specificity of 76.6%, surpassing the baseline 3D architecture (AUC = 0.55, sensitivity = 97.8%, specificity = 12.8%). Further analysis attributes the improved performance to the synergy between the generator model and discriminator architecture, whose robust feature representations effectively capture the intricate vasculature. Comparison with state-of-the-art 2D supervised en face classifiers demonstrates OCTA-GAN's ability to generalize across diverse artifacts and provides an interpretable organization of the output scores based on severity.