Deep Learning Application to Detect Glaucoma with a Mixed Training Approach: Public Database and Expert-Labeled...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Artificial intelligence (AI) shows promise for early ocular disease detection, but requires large, well-curated datasets.
• Acquiring extensive, high-quality medical image databases presents a significant challenge.
• This study employed a novel approach using public datasets for initial training and a specialized patient database for refinement.

Purpose of the Study:

• To develop and validate an AI-based glaucoma recognition algorithm.
• To assess the impact of retraining an algorithm with expert-labeled data on its diagnostic performance.
• To explore cost-effective methods for automated glaucoma screening.

Main Methods:

• A ResNet-50 deep learning architecture was adapted and initially trained on 10,658 public fundus images.
• An additional 1,158 expert-labeled images from 616 patients were incorporated, categorized by glaucoma severity.
• The algorithm underwent initial testing, followed by retraining on 70% of the specialized database and subsequent re-testing.

Main Results:

• Initial testing yielded an Area Under the Curve (AUC) of 76% for all images, with specific AUCs of 66% (early), 82% (moderate), and 84% (advanced) glaucoma.
• After retraining, the algorithm's AUC improved to 82% overall, with AUCs of 72% (early), 89% (moderate), and 91% (advanced) glaucoma.
• The combined data approach enhanced the system's precision in glaucoma detection.

Conclusions:

• Combining public and expert-curated datasets significantly improves AI algorithm precision for glaucoma identification.
• This strategy offers a promising pathway towards developing affordable tools for automated glaucoma screening.
• The refined algorithm demonstrates enhanced capability in detecting glaucoma at various disease stages.