Strategies to Improve Convolutional Neural Network Generalizability and Reference Standards for Glaucoma Detection...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma detection using artificial intelligence (AI) requires robust models.
• Convolutional neural networks (CNNs) show promise for detecting glaucoma from optical coherence tomography (OCT) images.
• Generalizability of AI models across different datasets and reference standards is a critical challenge.

Purpose of the Study:

• To enhance the generalizability of CNNs for glaucoma detection using OCT retinal nerve fiber layer probability maps and b-scans.
• To evaluate the impact of different reference standards (RS) on CNN accuracy for glaucoma detection.

Main Methods:

• CNNs were trained and evaluated on unseen datasets from different sites.
• Techniques to improve generalizability included data augmentation, multimodal input, and training with confidently rated images.
• Model performance was assessed using various RS.

Main Results:

• Data augmentation and training on confident images improved CNN accuracy by 5-9% on new datasets.
• Optimal CNN performance was achieved when training and testing used similar RS.
• The developed CNNs demonstrated robustness to variations in RS.

Conclusions:

• CNN generalizability for glaucoma detection can be enhanced through data augmentation, multimodal imaging, and confident image selection.
• Using a consistent RS for training and testing optimizes CNN generalization.
• These strategies are crucial for deploying reliable AI in glaucoma detection.