Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Diagnosing Anemia from Eye Videos: AI Reads Blood Vessel Signals
Researchers developed VesselNet, a deep learning model that predicts hemoglobin and red blood cell counts from brief magnified videos of eye blood vessels, offering noninvasive point-of-care blood testing.
Background
Traditional blood testing requires venipuncture. This study of 224 participants evaluated whether analyzing tiny eye blood vessels via 10-second videos could enable noninvasive blood biomarker assessment using AI.
Key Findings
- VesselNet achieved Spearman’s ρ = 0.47 for hemoglobin and ρ = 0.46 for RBC counts
- Anemia detection accuracy: 82.8% (95% CI: 73.9%-90.3%)
- Thin vessels (≤8 pixels) were more informative than thick vessels (ρ = 0.47 vs 0.30)
- 16-vessel analysis outperformed single-vessel approaches (AUC 82.8% vs 78.5%)
- Performance was robust across different populations
Why It Matters
This could serve as a triage tool for anemia screening in resource-limited or home-care settings, enabling frequent monitoring without venipuncture. The approach demonstrates scalability across diverse populations.
Limitations
Moderate correlations mean this cannot yet replace laboratory testing. Higher resolution imaging, larger cohorts, and improved algorithms are needed for clinical-grade accuracy.
Original paper: Towards noninvasive blood count using a deep learning pipeline from bulbar conjunctiva videos. — NPJ digital medicine. 10.1038/s41746-026-02598-2
