Equitable Deep Learning for Diabetic Retinopathy Detection Using Multidimensional Retinal Imaging With Fair Adaptive...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Ophthalmology

Background:

• Deep learning models are increasingly used for diabetic retinopathy (DR) detection.
• Existing models may exhibit performance disparities across different demographic groups.
• Ensuring fairness and equity in AI-driven medical diagnostics is crucial.

Purpose of the Study:

• To evaluate the fairness of current deep learning models for DR detection.
• To develop and introduce an equitable deep learning model to mitigate performance gaps.
• To improve the overall accuracy and fairness of DR detection systems.

Main Methods:

• Performance and fairness of deep learning models were assessed using fundus images and OCT B-scans.
• A novel Fair Adaptive Scaling (FAS) module was developed to address group disparities.
• Area Under the Receiver Operating Characteristic Curve (AUC) and equity-scaled AUC were used for evaluation.

Main Results:

• Integration of FAS with EfficientNet on fundus images improved overall AUC and equity-scaled AUC, with significant gains for Asian and White populations by race, and for gender.
• On OCT B-scans, FAS with DenseNet121 enhanced AUC and equity-scaled AUC, showing improvements for Asian and Black populations by race, and for both genders.
• The FAS module demonstrated a statistically significant improvement in both model performance and equity across different demographic groups.

Conclusions:

• Deep learning models for DR detection can exhibit performance variations across different demographic groups.
• The Fair Adaptive Scaling (FAS) module effectively enhances both the accuracy and equity of deep learning models for DR detection.
• The proposed equitable deep learning model shows potential for improving both performance and fairness in DR detection.