Sequential Deep Learning to Predict Non-Central to Central Geographic Atrophy Progression from OCT Imaging.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Geographic atrophy (GA) is a leading cause of vision loss in age-related macular degeneration (AMD).
• Predicting GA progression is crucial for timely therapeutic interventions.
• Current prediction methods often lack precision and multi-year forecasting capabilities.

Purpose of the Study:

• To develop and validate a temporal deep learning framework for predicting GA progression.
• Utilize longitudinal optical coherence tomography (OCT) sequences for multi-year GA forecasting.
• Enable automated, individualized risk stratification for GA management.

Main Methods:

• Retrospective analysis of 91 dry AMD patients with 455 OCT volumes.
• Feature extraction from OCT scans using ResNet and ViT architectures.
• Temporal modeling with RNN, LSTM, and Transformer networks for disease trajectory prediction.

Main Results:

• High accuracy (ROC-AUC 0.84-1.00) in predicting GA onset from no GA (NGA) over 2-6 years.
• Transformer models achieved peak AUC of 0.96 for predicting central GA (CGA) involvement.
• Longer input sequences and temporal interval encoding enhanced prediction performance and stability.

Conclusions:

• Temporal deep learning accurately predicts GA progression from longitudinal OCT data.
• The framework forecasts disease advancement across clinically relevant 2-6 year horizons.
• This technology supports automated risk stratification to guide complement inhibitor therapy in GA patients.