Effects of Study Population, Labeling and Training on Glaucoma Detection Using Deep Learning Algorithms.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma detection from fundus photographs is crucial for preventing vision loss.
• Deep learning algorithms offer potential for automated glaucoma screening.
• Evaluating algorithm performance across diverse datasets and training strategies is essential.

Purpose of the Study:

• To compare the performance of independently developed deep learning algorithms for glaucoma detection.
• To evaluate different strategies for incorporating new data into existing models.
• To assess the generalizability of deep learning models across varied populations.

Main Methods:

• Two independent deep learning models for glaucoma detection were developed at different institutions (UCSD and University of Tokyo).
• Models were trained and tested on fundus photograph datasets from diverse populations (Diagnostic Innovations in Glaucoma Study/African Descent and Glaucoma Evaluation Study, Matsue Red Cross Hospital).
• Performance was compared using three strategies: original models, sequential retraining, and combined training data.

Main Results:

• Original models showed similar performance on one dataset but significant differences on another, indicating variability.
• Model performance was higher for moderate-to-severe glaucoma compared to mild cases.
• The combined training strategy generally yielded better performance across all tested datasets compared to the original models.

Conclusions:

• Deep learning models can achieve high accuracy in glaucoma detection across diverse datasets when employing appropriate training strategies.
• Model performance is influenced by disease severity, labeling, training methods, and population characteristics.
• Stratified reporting of accuracy by relevant covariates is crucial for reliable cross-study comparisons.