Automated OCT angiography image quality assessment using a deep learning algorithm.

PURPOSE: To expedite and to standardize the process of image quality assessment in optical coherence tomography angiography (OCTA) using a specialized deep learning algorithm (DLA).
METHODS: Two hundred randomly chosen en-face macular OCTA images of the central 3 × 3 mm2 superficial vascular plexus were evaluated retrospectively by an OCTA experienced reader. Images were defined either as sufficient (group 1, n = 100) or insufficient image quality (group 2, n = 100) based on Motion Artifact Score (MAS) and Segmentation Accuracy Score (SAS). Subsequently, a pre-trained multi-layer deep convolutional neural network (DCNN) was trained and validated with 160 of these en-face OCTA scans (group 1: 80; group 2: 80). Training accuracy, validation accuracy, and cross-entropy were computed. The DLA was tested in detecting 40 untrained OCTA images (group 1: 20; group 2: 20). An insufficient image quality probability score (IPS) and a sufficient image quality probability score (SPS) were calculated.
RESULTS: Training accuracy was 97%, validation accuracy 100%, and cross entropy 0.12. A total of 90% (18/20) of the OCTA images with insufficient image quality and 90% (18/20) with sufficient image quality were correctly classified by the DLA. Mean IPS was 0.88 ± 0.21, and mean SPS was 0.84 ± 0.19. Discrimination between both groups was highly significant (p < 0.001). Sensitivity of the DLA was 90.0%, specificity 90.0%, and accuracy 90.0%. Coefficients of variation were 0.96 ± 1.9% (insufficient quality) and 1.14 ± 1.6% (sufficient quality).
CONCLUSIONS: Deep learning (DL) appears to be a potential approach to automatically distinguish between sufficient and insufficient OCTA image quality. DL may contribute to establish image quality standards in this recent imaging modality.