Widen the Applicability of a Convolutional Neural-Network-Assisted Glaucoma Detection Algorithm of Limited Training...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Automated glaucoma detection using deep learning can enhance diagnostic rates and prevent blindness.
• Generalizable deep learning models for glaucoma detection are currently unavailable, even with large datasets.

Purpose of the Study:

• To evaluate a convolutional neural network (CNN) classifier's performance in glaucoma detection using limited high-quality fundus images.
• To identify methods for improving the CNN classifier's performance across diverse datasets.

Main Methods:

• A CNN classifier (EfficientNet B3) was trained on 944 fundus images from one medical center (core model).
• The core model was externally validated on three independent datasets.
• Performance was compared against an integrated model (all datasets) and dataset-specific models (fine-tuned).

Main Results:

• The core model achieved 95.62% accuracy but performance dropped to 52.5-80.0% on external datasets.
• Dataset-specific models demonstrated superior performance, achieving 87.50-92.5% diagnostic accuracy on external datasets.
• Fine-tuning improved model applicability across datasets, outperforming models trained on larger, integrated datasets.

Conclusions:

• Dataset-specific tuning of CNN classifiers is an effective strategy to enhance glaucoma detection performance across different datasets.
• This approach improves generalizability when simply increasing training data does not suffice.
• Fine-tuning offers a practical solution for deploying reliable AI in diverse clinical settings for glaucoma screening.