Enhancing Glaucoma Detection through Supervised Pre-training with Intermediate Phenotypes: A Multi-Institutional Study

Glaucoma is a leading cause of irreversible blindness worldwide, with early diagnosis often hindered by subtle symptomatology and the lack of comprehensive screening programs. In this study, we introduce a robust deep learning framework that leverages supervised pre-training with clinically relevant intermediate indicators, most notably the vertical cup-to-disc ratio (VCDR), to enhance glaucoma detection from color fundus images. Utilizing the expansive AIROGS dataset for pre-training, our multi-task learning strategy simultaneously addresses categorical diagnostic classification and VCDR regression. We evaluated three architectures, ResNet-18, DINOv2, and RETFound across multiple international cohorts, including DRISHTI, G1020, ORIGA, PAPILA, REFUGE1, and ACRIMA. Compared to out-of-domain pre-training or self-supervised pre-training, supervised pre-training achieved the best average performances on all of ResNet-18 (average AUROC = 0.857), DINOv2 (average AUROC = 0.788), and RETFound (average AUROC = 0.839). The best model is the ResNet-18 pre-trained with AIROGS diagnostic features, achieving the highest AUROC of 0.930 on the G1020 dataset, and an average AUROC of 0.857 across all datasets. These results underscore the superiority of incorporating domain-specific clinical labels to guide feature extraction, thereby improving model performance, and generalizability for glaucoma detection. Our findings advocate for the integration of supervised pre-training strategies into glaucoma detection model development, with significant potential to improve model performance and ultimately improve patient outcomes with decreased undiagnosed rate.