Deep Learning Approaches to Predict Geographic Atrophy Progression Using Three-Dimensional OCT Imaging.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Age-related macular degeneration (AMD) is a leading cause of vision loss.
• Geographic atrophy (GA) is an advanced form of AMD characterized by progressive retinal degeneration.
• Accurate prediction of GA progression is crucial for monitoring disease and evaluating treatments.

Purpose of the Study:

• To assess the performance of different processing methods for 3D optical coherence tomography (OCT) images.
• To determine the efficacy of these methods when used with deep learning models for predicting GA lesion area and growth rate.
• To identify optimal image processing strategies for GA progression prediction using OCT data.

Main Methods:

• Utilized OCT volumes from lampalizumab clinical trials, including 1219 eyes for development and 442 for evaluation.
• Evaluated four OCT image processing approaches: en-face intensity maps, SLIVER-net, 3D convolutional neural networks (CNNs), and segmentation-derived maps.
• Employed CNN models to predict baseline GA lesion size and annualized growth rate using processed OCT data.

Main Results:

• All evaluated approaches showed comparable performance in predicting GA growth rate (r2 ≈ 0.33–0.35).
• Baseline GA lesion size prediction was also comparable (r2 ≈ 0.9–0.91), with SLIVER-net performing slightly lower (r2 = 0.83).
• Thickness maps of the ellipsoid zone (EZ) and retinal pigment epithelium (RPE) layers were most informative; combining them improved predictions, while adding other layers did not.

Conclusions:

• Current processing approaches achieve comparable, potentially plateaued, performance in predicting GA growth rate.
• The EZ and RPE retinal layers contain the most significant information for predicting GA progression.
• 3D OCT imaging holds substantial utility for predicting GA disease progression.