Quantification of nonperfusion area in montaged wide-field optical coherence tomography angiography using deep learning in diabetic retinopathy

Abstract Purpose To examine the efficacy of a deep-learning-based algorithm to quantify the nonperfusion area (NPA) on montaged wide-field OCTA for assessment of DR severity Design Cross-sectional study. Participants A total of 137 participants with a full range of DR severity and 26 normal participants. Methods A deep-learning-based algorithm was developed for detecting and quantifying NPA in the superficial vascular complex on wide-field OCTA composed of three horizontally montaged 6x6-mm OCTA scans from the nasal, macular, and temporal regions. We trained the algorithm on 978 volumetric OCTA scans from all participants using 5-fold cross-validation. The algorithm can distinguish NPA from shadow artifacts. The F1-score evaluated segmentation accuracy. The area under the receiver-operating characteristic curve (AROC) and sensitivity with specificity fixed at 95% quantified network performance to distinguish diabetics from healthy controls, referable DR from non-referable DR (non-proliferative DR below moderate severity), and severe DR (severe non-proliferative DR, proliferative DR, or DR with edema) from non-severe DR (mild to moderate non-proliferative DR). Main outcome measures Wide-field OCTA NPA, visual acuity and DR severities. Results Automatically segmented NPA had a high agreement with the manually delineated ground truth, with an F1-score of 0.78±0.05 (mean ± standard deviation) in nasal, 0.82±0.07 in macular, and 0.78±0.05 in temporal scans. The extrafoveal avascular area (EAA) in the macular scan had the best sensitivity at 54% for differentiating diabetics from healthy controls, while the montaged wide-field OCTA scan had a significantly higher sensitivity than macular scans (p Conclusions A deep-learning-based algorithm on montaged wide-field OCTA can detect NPA accurately and improve the detection of clinically important DR.