Deep learning for retinal non-perfusion and foveal avascular zone analysis in wide-field OCTA in diabetic retinopathy.

We developed an automated framework for segmenting low-quality and non-perfusion areas in widefield OCTA images to obtain two key metrics useful for diabetic retinopathy (DR) monitoring: the retinal non-perfusion index (NPI) and foveal avascular zone (FAZ) area. Using 170 images from 88 patients in the EVIRED cohort, we trained two models: Q-NET, which segments low-quality areas, and NPA-NET, which detects non-perfusion areas and the FAZ. Their combined outputs created a 4-class map to calculate NPI and FAZ area. Ground truth segmentations were established by a single expert (for non-perfusion and FAZ areas) or a consensus of four annotators (for low-quality areas). NPA-NET and Q-NET, tested on 29 images, achieved strong segmentation performances (Dice coefficients of 0.714 (low-quality), 0.781 (non-perfusion), and 0.879 (FAZ)). Some inter-annotator variability was found (mean Dice: 0.85 for low-quality, 0.683 for non-perfusion areas). Predictive accuracy for NPI and FAZ area was high, with R² coefficients of 0.97 and 0.63, respectively, with minimal underestimation and no overestimation. This AI tool provides reliable biomarkers for DR monitoring, supporting treatment decisions and medical decision-making by automatically analyzing OCTA images, and could be integrated into clinical practice.