Equity-enhanced glaucoma progression prediction from OCT with knowledge distillation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a progressive optic neuropathy leading to irreversible vision loss.
• Accurate glaucoma progression prediction is essential for timely and effective treatment strategies.
• Deep learning models show promise for progression prediction but can exhibit demographic biases.

Purpose of the Study:

• To develop and evaluate FairDist, a novel deep learning model for predicting glaucoma progression.
• To ensure equitable performance of the model across different gender and racial groups.
• To compare FairDist against existing methods in terms of accuracy and fairness.

Main Methods:

• Utilized baseline optical coherence tomography (OCT) scans for glaucoma progression prediction.
• Employed an equity-aware EfficientNet for initial glaucoma detection, adapted via knowledge distillation for progression prediction.
• Assessed model performance using Area Under the Curve (AUC), Sensitivity, and Specificity.
• Quantified fairness using equity-scaled AUC, adjusting for subgroup disparities.

Main Results:

• FairDist achieved superior AUC and equity-scaled AUC compared to baseline and fairness-aware methods.
• The model demonstrated high accuracy and fairness across both gender and racial subgroups.
• Performance was evaluated for mean deviation, fast progression, and total deviation pointwise progression.

Conclusions:

• FairDist offers a robust and equitable approach to predicting glaucoma progression.
• The model's fairness mitigation strategies are effective in reducing demographic disparities.
• FairDist has the potential for broader application in predicting other disease progression tasks with improved performance and fairness.