Deep Learning Model Stages AMD with 85% Agreement Using OCT/OCTA Scans

Researchers led by Yukun Guo developed and evaluated three deep learning models for automated staging of age-related macular degeneration (AMD) severity using optical coherence tomography (OCT) and OCT angiography (OCTA) data. The study analyzed 2,030 OCT/OCTA volumes from 351 eyes of 271 participants aged 50 years and older with varying AMD severities. AMD severity was graded into four stages (No AMD, Early AMD, Intermediate AMD, Advanced AMD) according to the AREDS simplified severity scale. Three EfficientNet-based deep learning models were trained using different input modalities: (1) biomarker maps derived from segmented pathological features (retinal fluid, drusen, geographic atrophy, macular neovascularization); (2) 2D en face OCT and OCTA projections; and (3) 3D OCT/OCTA volumes. All models used normalized inputs, data augmentation, and five-fold cross-validation.

All models demonstrated strong staging performance with substantial agreement with the reference standard (QWK >= 0.83). The biomarker-based model achieved the highest overall performance (QWK = 0.85 ± 0.03) and best detection of early AMD (F1-score = 0.59 ± 0.14). The 2D OCT/OCTA model showed the highest precision (0.79 ± 0.06) and most accurately identified eyes without AMD, while the 3D model performed comparably to the 2D model (QWK = 0.83 ± 0.04 vs. 0.83 ± 0.09). These results indicate that deep learning models using OCT/OCTA data can accurately and automatically grade AMD severity, with the biomarker-based model providing the most balanced performance and particular value for early AMD detection.