Predicting Age-related Macular Degeneration Progression with Longitudinal Fundus Images Using Deep Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Predicting progression to late age-related macular degeneration (AMD) is vital for personalized treatment.
• Current risk models often lack the ability to leverage historical patient data, specifically longitudinal color fundus photographs (CFPs).

Purpose of the Study:

• To evaluate deep neural networks' capability in utilizing sequential CFP data for improved prediction of late AMD risk.
• To assess the efficacy of novel deep learning architectures in capturing temporal information from longitudinal CFPs.

Main Methods:

• Proposed two deep learning models: CNN-LSTM and CNN-Transformer, integrating Convolutional Neural Networks (CNNs) with Long-Short Term Memory (LSTM) and Transformer architectures, respectively.
• Evaluated models on the Age-Related Eye Disease Study (AREDS) cohort using longitudinal CFPs.
• Compared performance against baseline models using single-visit CFPs.

Main Results:

• The proposed CNN-LSTM and CNN-Transformer models demonstrated superior performance in predicting 2-year and 5-year risks of progression to late AMD compared to single-visit baselines.
• Achieved higher Area Under the Curve (AUC) values: 0.879 vs 0.868 for 2-year, and 0.879 vs 0.862 for 5-year predictions.
• Longer time intervals of longitudinal CFP data positively impacted prediction accuracy.

Conclusions:

• Deep learning models effectively capture sequential information from longitudinal CFPs for enhanced AMD progression risk prediction.
• Utilizing historical imaging data significantly improves the accuracy of predicting future late AMD.
• The developed models and open-source code can advance research in AI-driven ophthalmic diagnostics.