Deep Learning Assessment of Intraretinal Microvascular Abnormalities in Diabetic Retinopathy on Swept-Source Optical Coherence Tomography Angiography

Purpose To develop and evaluate a deep learning algorithm for automated segmentation and quantification of intraretinal microvascular abnormalities (IRMAs) on swept-source optical coherence tomography angiography (SS-OCTA) across multiple fields of view (FOVs) and explore IRMA burden as an optical coherence tomography angiography (OCTA) biomarker for stratifying diabetic retinopathy (DR) severity and correlation with retinal nonperfusion. Design Cross-sectional study. Subjects 254 eyes from 181 patients with varying stages of diabetic retinopathy, including non-proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy (PDR). Methods Eyes of different DR stages were imaged using 6×6mm, 12×12mm, and 29×24mm scans. A context-enhanced U-Net (CE-U-Net) incorporating nonperfusion area (NPA) priors and attention-based feature recalibration was trained to segment IRMAs on en face all-retinal-layer (ARL) OCTA images. Main Outcome Measures Performance was evaluated using lesion-level F1 score, intraclass correlation coefficients (ICC) for quantitative agreement with ground-truth IRMA counts, and Bland–Altman analysis. IRMA counts were compared across DR severity stages and correlated with nonperfusion index (NPI). Results A total of 254 OCTA scans were analyzed (NPDR = 60.2%, PDR=39.8%). The CE-U-Net achieved mean F1 scores of 0.61±0.08, 0.71±0.02, and 0.65±0.05 for IRMA segmentation on 6×6mm, 12×12mm, and 29×24mm scans, respectively. Algorithm-predicted IRMA counts showed strong agreement with ground-truth across all FOVs (ICC=0.81, 0.86, 0.82 respectively). OCTA-detectable IRMAs were present from mild-moderate NPDR and increased markedly from mild-moderate to severe NPDR, particularly in 29×24mm scans (p<0.001), before plateauing in PDR (p=0.72). The strongest correlation between IRMA counts and NPI was observed in 12×12mm scans across all DR stages (p=0.01, <0.001 and <0.001 respectively). Conclusion Automated IRMA segmentation and quantification on SS-OCTA is feasible and reproducible across multiple FOVs, including UWF scans. By reliably capturing both clinically visible and OCTA-detectable "occult" IRMAs, this approach supports IRMA burden as a quantitative biomarker for stratifying DR severity and advances the case for biomarker-driven DR assessment.