Equitable deep learning for diabetic retinopathy detection using multi-dimensional retinal imaging with fair adaptive scaling: a retrospective study

Summary Background As deep learning becomes increasingly accessible for automated detection of diabetic retinopathy (DR), questions persist regarding its performance equity among diverse identity groups. We aimed to explore the fairness of current deep learning models and further create a more equitable model designed to minimize disparities in performance across groups. Methods This study used one proprietary and two publicly available datasets, containing two-dimensional (2D) wide-angle color fundus images, scanning laser ophthalmoscopy (SLO) fundus images, and three-dimensional (3D) Optical Coherence Tomography (OCT) B-Scans, to assess deep learning models for DR detection. We developed a Fair Adaptive Scaling (FAS) module that dynamically adjusts the significance of samples during model training for DR detection, aiming to lessen performance disparities across varied identity groups. FAS was incorporated into both 2D and 3D deep learning models to facilitate the binary classification of DR and non-DR cases. The area under the receiver operating characteristic curve (AUC) was adopted to measure the model performance. Additionally, we devised an equity-scaled AUC metric, which evaluates model fairness by balancing overall AUC against disparities among groups. Findings Using in-house color fundus images on the racial attribute, the overall AUC and ES-AUC of EfficientNet, after integrating with FAS, improved from 0.88 and 0.83 to 0.90 and 0.84 (p < 0.05), with AUCs for Asians and Whites improving by 0.04 and 0.03, respectively (p < 0.01). Regarding gender, both the overall AUC and ES-AUC of EfficientNet improved by 0.01 (p < 0.05) after integrating with FAS. While using in-house SLO fundus images based on race, the overall AUC and ES-AUC of EfficientNet improved from 0.80 to 0.83 (p < 0.01), with AUCs for Asians, Blacks, and Whites improving by 0.02, 0.01 and 0.04, respectively (p < 0.05). On gender, FAS improved the overall AUC and ES-AUC of EfficientNet by 0.02, with both genders showing an improvement of 0.02 (p < 0.01). Using the 3D deep learning model DenseNet121 on in-house OCT-B-Scans based on race, FAS improved the overall AUC and ES-AUC from 0.875 and 0.81 to 0.884 and 0.82 respectively, where the AUCs for Asians and Blacks improved by 0.03 and 0.02 (p < 0.01). On gender, FAS improved the overall AUC and ES-AUC of DenseNet121 by 0.04 and 0.03, while the AUCs for Females and Males improved by 0.05 and 0.04 (p < 0.01), respectively. Interpretation Existing deep learning models indeed exhibit variable performance across diverse identity groups in DR detection. The FAS proves beneficial in enhancing model equity and boosting DR detection accuracy, particularly for underrepresented groups.