A Deep-Learning Algorithm to Predict Short-Term Progression to Geographic Atrophy on Spectral-Domain Optical...

Area of Science:

• Ophthalmology and Artificial Intelligence
• Medical Imaging Analysis
• Retinal Disease Progression

Background:

• Identifying patients at risk of geographic atrophy (GA) progression from intermediate age-related macular degeneration (iAMD) is crucial for developing preventative clinical trials.
• Current methods for predicting GA progression can be improved for greater accuracy and efficiency.

Purpose of the Study:

• To develop and validate a deep learning algorithm, DeepGAze, for predicting the progression of iAMD to GA within one year.
• To assess the algorithm's performance using volumetric spectral-domain optical coherence tomography (SD-OCT) scans.

Main Methods:

• A retrospective cohort study utilized SD-OCT scans from multiple datasets (AREDS2, clinical care).
• A position-aware convolutional neural network was trained and validated on these datasets.
• Performance was evaluated using metrics including AUROC, AUPRC, sensitivity, and specificity.

Main Results:

• The DeepGAze algorithm demonstrated high accuracy in predicting GA progression, with an AUROC of 0.94 on the primary dataset and 0.94 on an external validation dataset.
• The fully automated model performed comparably to models incorporating expert-annotated features.
• Simulated clinical trial recruitment showed significant enrichment for patients progressing to GA.

Conclusions:

• The fully automated DeepGAze algorithm accurately predicts iAMD to GA progression within a clinically relevant timeframe.
• This predictive capability can significantly enhance the design and efficiency of clinical trials for GA prevention.
• The algorithm holds potential for guiding clinical decisions regarding patient monitoring and treatment initiation.