Strong versus Weak Data Labeling for Artificial Intelligence Algorithms in the Measurement of Geographic Atrophy.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) measurement is crucial for monitoring age-related macular degeneration.
• Accurate GA quantification requires detailed image analysis, often involving manual segmentation.
• Deep learning models show promise for automating GA measurement from fundus autofluorescence (FAF) images.

Purpose of the Study:

• To evaluate data labeling requirements for training deep learning models for GA measurement using FAF images.
• To compare the performance of AI models trained with different levels of data labeling.

Main Methods:

• Two sets of FAF images were used: weakly labeled (area measurements only) and strongly labeled (GA segmentation masks).
• The Age-Related Eye Disease Study 2 (AREDS2) dataset was used for training and cross-validation.
• Clinical trial images were used for testing AI models.

Main Results:

• Deep learning models achieved comparable GA area measurements to human graders, even with weakly labeled data.
• The Dice coefficient demonstrated high accuracy for segmentation, with values of 0.89 for cross-validation and 0.92 for testing.
• Integrating large volumes of weakly labeled images with a small number of strongly labeled images proved effective.

Conclusions:

• Deep learning models can achieve reasonable accuracy in GA measurement using weakly labeled FAF images.
• Hybrid training approaches combining weakly and strongly labeled data are a cost-effective solution for AI development in ophthalmology.