Computer vision for sclera segmentation and non-invasive jaundice assessment from smartphone eye images (Pfizer CDI).
Computer vision models for automatic sclera segmentation from smartphone-captured eye images, with edge-deployed pipelines for on-device, real-time inference. In collaboration with the University General Hospital of Patras, the work extends to statistical and causal modeling of digital biomarkers and hematological indicators for non-invasive jaundice assessment, under institutional ethics approval.
Developed under the externally funded Pfizer CDI — EyeAI research contract at the University of Patras.
We provide the first systematic, PRISMA-guided survey of AI-based methods for non-invasive jaundice detection and bilirubin prediction from adult ocular images. Following PRISMA guidelines, 47 studies were included from 520 retrieved records. A five-axis taxonomy categorises studies by task type, AI methodology, sclera segmentation strategy, input acquisition and calibration, and deployment setting. The best-performing bilirubin regression systems achieved Pearson correlations of 0.89 and R\textsuperscript2 of 0.956; state-of-the-art sclera segmentation reached F1-scores of 96.66% and IoU of 93.59%. Five critical gaps were identified: small datasets, absent standardised evaluation protocols, dominance of binary classification, disconnect between sclera segmentation and jaundice detection communities, and absence of prospective clinical validation.
@article{kontogiannis2026jaundice,author={Kontogiannis, George and Tzamalis, Pantelis and Velissaris, Dimitrios and Nikoletseas, Sotiris},title={{AI}-Based Jaundice Assessment from Ocular Images: A Structured Survey of Methods, Datasets, and Open Challenges},journal={Healthcare Informatics Research},year={2026},note={Under Review},}