Deep Learning to Predict Geographic Atrophy Area and Growth Rate from Multimodal Imaging.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) is a leading cause of vision loss.
• Accurate prediction of GA growth rate is crucial for clinical trial design and patient management.
• Current methods for assessing GA progression can be time-consuming and subjective.

Purpose of the Study:

• To develop and validate deep learning models for predicting the annualized geographic atrophy (GA) growth rate.
• To utilize fundus autofluorescence (FAF) images and spectral-domain optical coherence tomography (OCT) volumes for GA growth prediction.
• To assess the potential of these models for prognostic covariate adjustment in clinical trials.

Main Methods:

• Retrospective analysis of GA growth rate estimation using linear fit of lesion area over two years.
• Development of three multitask deep learning models: FAF-only, OCT-only, and multimodal (FAF + OCT).
• Training and testing models on development, holdout, and independent test datasets from multiple clinical trials.

Main Results:

• Multitask deep learning models demonstrated high accuracy in predicting GA lesion area (r² up to 0.98) and growth rate (r² up to 0.65) on independent test sets.
• The FAF-only model showed strong performance for both lesion area and growth rate prediction.
• Model performance was evaluated using squared Pearson correlation coefficient (r²) with confidence intervals.

Conclusions:

• Deep learning models can effectively predict individual GA area and growth rates from baseline FAF and OCT imaging.
• The developed models show feasibility for use in prognostic covariate adjustment to improve clinical trial power.
• This approach offers a promising tool for advancing research and treatment strategies for geographic atrophy.