Image Severity Ranking in Ophthalmology
Expert-centric severity representation learning for medical images
Sep 2022 – Dec 2023 · USC, with collaborators in ophthalmology
Clinicians reason about disease as a matter of degree, not just presence or absence. This project learns representations of disease severity from medical images, with a focus on glaucoma. The goal is models that rank images the way experts do and can explain their rankings.
Highlights
- ConPro: a severity representation learning framework combining contrastive learning and preference optimization. It outperforms state-of-the-art methods by 6% in F1 score (Nguyen et al., 2024).
- SiameN and SEMISE: expert-centric severity ranking models, through explainable pairwise n-hidden comparison (Nguyen et al., 2025) and semi-supervised severity representation learning (Tran et al., 2025). Together they improve IoU explainability by 6% and ranking performance by 3% in F1 score.
- Clinical evaluation: comparing deep learning and clinician performance for detecting referable glaucoma from fundus photographs in a safety-net population (Nguyen et al., 2025).
References
2025
- OphthSciComparison of Deep Learning and Clinician Performance for Detecting Referable Glaucoma from Fundus Photographs in a Safety Net PopulationOphthalmology Science, 2025
