Topographic Clinical Insights From Deep Learning-Based Geographic Atrophy Progression Prediction

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) is a leading cause of vision loss.
• Predicting GA growth rate is crucial for clinical management.
• Fundus autofluorescence (FAF) imaging provides topographic information relevant to GA progression.

Purpose of the Study:

• To determine the contribution of specific FAF topographic imaging features to the performance of deep learning (DL) algorithms in predicting GA growth rate.
• To identify which FAF image regions are most informative for DL-based GA progression prediction.

Main Methods:

• Retrospective analysis of FAF images from GA clinical trials.
• Development and benchmark training of a DL algorithm using full FAF images.
• Ablation experiments systematically removing or shuffling pixels from specific FAF regions (Lesion, Rim, Background).
• Evaluation of algorithm performance using Squared Pearson correlation (r2) to compare ablated datasets against the benchmark.

Main Results:

• The Rim region demonstrated the highest influence on predictive performance (r2) compared to Lesion and Background regions.
• Shuffling pixels or using masks for any region resulted in similar performance, highlighting the importance of textural context over simple intensity.
• Lesion size, evaluated via a Convex Hull dataset, also contributed to predictive accuracy.

Conclusions:

• Topographic FAF imaging features, particularly the Rim region, significantly contribute to the accuracy of DL algorithms for GA growth rate prediction.
• Textural information within FAF images is more critical than isolated intensity values for predicting GA progression.
• These findings offer valuable clinical insights for developing more effective AI-driven tools for GA management.