Quantifying Geographic Atrophy in Age-Related Macular Degeneration: A Comparative Analysis Across 12 Deep Learning...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) is a leading cause of vision loss.
• Accurate segmentation of GA from fundus autofluorescence (FAF) images is crucial for monitoring disease progression and evaluating treatments.
• Artificial intelligence (AI) offers promising tools for automated GA segmentation.

Purpose of the Study:

• To evaluate the effectiveness of 12 different AI architecture combinations for segmenting geographic atrophy (GA) from fundus autofluorescence (FAF) images.
• To identify the optimal AI architecture for accurate and reliable GA segmentation.

Main Methods:

• Investigated combinations of AI encoders (EfficientNet, ResNet, VGG, mViT) and decoders (FPN, UNet, PSPNet).
• Trained models on 601 FAF images from the AREDS2 study and validated on 156 images from the GlaxoSmithKline study.
• Performance was assessed by comparing AI-predicted GA areas to human measurements and calculating the Dice Coefficient (DC).

Main Results:

• The best-performing AI models for GA segmentation were UNet and FPN frameworks utilizing the Mix Vision Transformer (mViT) encoder.
• The Dice Coefficient (DC) ranged from 0.884 to 0.993 across different architectures.
• PSPNet frameworks demonstrated the least effective performance in GA segmentation.

Conclusions:

• AI architecture selection is a critical factor influencing the performance of GA segmentation from FAF images.
• Vision transformer-based models, particularly when combined with FPN and UNet architectures, show superior suitability compared to traditional CNN-based models.
• Tailoring AI architecture choice to specific project goals is essential for optimal results in GA segmentation.