Automated quantification of parafoveal microvascular complexity using deep learning-based OCTA segmentation and foveal avascular zone detection.

BACKGROUND: Quantitative analysis of retinal microvasculature from optical coherence tomography angiography (OCTA) is limited by artifacts and manual segmentation variability. This study aimed to develop and validate a fully automated deep learning-based pipeline for OCTA segmentation and parafoveal microvascular complexity quantification. METHODS: This cross-sectional image-based study used 6 × 6 mm macular OCTA scans (512 × 512 pixels) obtained with the Optovue AngioVue system. A total of 132 eyes from 132 patients were included, with one eye selected randomly per patient to ensure statistical independence. A U-Net convolutional neural network (CNN) was trained on manually labeled vessel masks for vascular segmentation. Skeleton-based features-including branch count, segment count, and branch density index (BDI)-were extracted from the binarized masks. The largest avascular component was automatically detected as the foveal avascular zone (FAZ), and the parafoveal region (500 μm radius) was analyzed. RESULTS: The deep learning model achieved Dice coefficient = 0.946 ± 0.012, IoU = 0.898 ± 0.021, precision = 0.927 ± 0.015, and recall = 0.967 ± 0.009 on the independent test set (n = 27), confirming robust generalizability to unseen data. The dedicated FAZ segmentation model achieved Dice = 0.648 ± 0.116 on its independent test set (n = 9). The full analysis for each image completed in < 2 s using a single GPU. CONCLUSIONS: The proposed deep learning framework enables rapid, objective, and reproducible quantification of retinal microvascular complexity. Its integration into OCTA software may facilitate large-scale clinical screening and quantitative monitoring of retinal vascular health. CLINICAL TRIAL NUMBER: Not applicable.