Abstract
Purpose :
Eye movements, optical opacities, and a variety of other factors [1] can introduce artifacts during the acquisition of optical coherence tomography angiography (OCTA) volumes. We aim to develop an unsupervised, automated deep learning model to separate poor-quality from good-quality volumes in a quantitative and objective manner.
Methods :
This is a cross-sectional and observational study of human subject data collected under an IRB approved protocol. OCTA 3×3mm volumes were acquired using an Angioplex (Zeiss Cirrus 5000 Spectral Domain) and were labeled by trained graders as “Good” (N=451, 405 train, 46 test), “Questionable” (N=46), and “Poor” (N=46) based on severity of artifacts including media opacity, image focus, centration, and movement. OCTA-GAN, a novel Generative Adversarial Network (GAN) [2] architecture, was developed that incorporates multi-scale layer processing to detect fine-grained artifacts by fusing vasculature details with larger anatomical context. A novel generator up-sampling layer improved the model’s parametric and computational efficiency. The model was trained exclusively on “Good” volumes in PyTorch with a generator learning rate (LR) of 0.0001, a discriminator LR of 0.0002, and an Adam optimizer. Class activation maps illustrated the improvements of the OCTA-GAN architecture. An ANOVA and t-test were used to make comparisons.
Results :
The discriminator achieved an AUC of 0.90 separating “Good” from “Poor” (p<0.001) surpassing the α-GAN baseline architecture (AUC = 0.57). A discriminator score threshold of 3.42 yields 97.8% sensitivity and 71.7% specificity for “Good” volumes. The “Questionable” discriminator score distribution was between the “Good” and “Poor” distributions, consistent with the graded labels (ANOVA p<0.001). Grad-CAM activations show OCTA-GAN focuses on relevant OCTA vasculature while α-GAN does not. The OCTA-GAN generator’s efficient up-sampling layers required only 75G floating point operations compared to 224G for standard layers. OCTA-GAN’s discriminator (21M) requires fewer parameters than α-GAN (28M).
Conclusions :
We demonstrate our OCTA-GAN model accurately distinguishes between “Good” and “Poor” quality OCTA volumes with a high sensitivity. Experimental results validate OCTA-GAN’s improved focus on fine microvasculature and reduced model complexity over standard architectures.
[1] Spaide, Retina, 2015, [2] Arjovsky, PMLR, 2017
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.