Developing Equitable Deep Learning Models for Automated Glaucoma Screening - Project Summary
Motivation: Glaucoma is a leading cause of irreversible blindness affecting 3 million patients in the US. About
50% of people with glaucoma are unaware of their condition, and racial and ethnic minority groups are
particularly affected due to a lack of access to affordable ophthalmic care. Glaucoma detection with deep
learning using retinal imaging is promising to provide affordable glaucoma screening to reduce health
disparities, which can be deployed in primary care and pharmacies without needing individuals to visit the more
expensive eye clinics for screening. It remains unclear if existing deep learning models for glaucoma detection
are equitable, and minimal studies have been conducted to develop novel methods to reduce performance
disparities. We hypothesize that more equitable performance in glaucoma detection can be achieved by
developing equity-aware deep learning techniques. We propose new equity-scaled performance metrics of the
area under the receiver operating characteristic curve, sensitivity, and specificity for model equity assessment,
which penalize raw performance metrics with performance variability across identity groups.
Overall Approach: Our models use optical coherence tomography scans and fundus photos to predict
glaucoma diagnosis defined by clinical guidelines. We use six identity attributes: gender, race, ethnicity,
preferred language, marital status, and income level. Our equity-aware models will be compared with standard
models with and without oversampling and transfer learning measured by identity-group specific and equity-
scaled overall accuracy metrics. Aims: (1) Assessing performance equity of existing deep learning
models and developing a fair identity scaling approach. We will quantify the identity-group-specific
performance for the three widely used deep learning models including ResNet, EfficientNet, and Vision
Transformer. We will develop a novel fair identity scaling approach to equalize performance across identity
groups by reweighting the training target function by the error magnitude of each identity group proportionally.
(2) Developing identity-conditioned generative models to improve data equality for improving
performance equity. We will develop identity-conditioned generative models to synthesize identity-diagnosis-
specific training data. The generated data will be used to equalize the data availability across all identity groups
for training detection models. (3) Developing fair identity normalization models to improve glaucoma
detection equity. We will develop a fair identity normalization technique to equalize the feature importance
across identity groups. We will integrate the three equity-aware approaches to maximize performance equity.
Main Deliverables and Impacts: Our equity-aware deep learning models for automated glaucoma screening
will greatly benefit racial and ethnic minorities and socioeconomically disadvantaged groups. We will release a
large de-identified dataset with 10,000 patients to the public to study medical artificial intelligence equity.
