Deep Learning for Prediction of AMD Progression: A Pilot Study.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Age-related macular degeneration (AMD) is a leading cause of vision loss.
• Predicting the conversion from early/intermediate to advanced wet AMD is crucial for timely intervention.
• Optical coherence tomography (OCT) provides detailed cross-sectional retinal images.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for predicting wet AMD progression.
• To assess the utility of OCT imaging in conjunction with deep convolutional neural networks (CNNs).
• To investigate the impact of preprocessing techniques on prediction accuracy.

Main Methods:

• Seventy-one eyes with early/intermediate AMD were imaged using OCT over two years.
• Two CNN architectures (VGG16 and a novel AMDnet) were trained and validated.
• Segmentation-based normalization preprocessing was applied to OCT data.

Main Results:

• The novel AMDnet architecture with preprocessing achieved an AUC of 0.89 (B-scan) and 0.91 (volume).
• VGG16 with preprocessing yielded AUCs of 0.82 (B-scan) and 0.87 (volume).
• Preprocessing significantly improved performance for both CNN architectures.

Conclusions:

• A CNN incorporating layer segmentation-based preprocessing demonstrates high predictive power for AMD progression.
• The developed method shows promise for identifying patients at risk of converting to advanced wet AMD.
• Preprocessing is essential for enhancing the performance of deep learning models in OCT-based AMD prediction.