Most anemia screening tools that skip the needle still can't beat a basic blood draw for actually measuring hemoglobin levels - binary "anemic or not" classifiers hit 97%+ accuracy, but ask them to predict a continuous hemoglobin value and they start sweating. A new paper in NPJ Digital Medicine takes a different angle entirely: instead of snapping a photo of your eye, they film it - and the results suggest your conjunctival capillaries have been narrating your blood story this whole time.

Wait, You Can Read Blood Through Someone's Eye?

Doctors have eyeballed (pun intended) the inside of eyelids for anemia signs since forever. Pull down the lower lid, check if it looks pale - boom, clinical assessment. The problem? That bedside trick has roughly 18-25% sensitivity (Sheth et al., PLOS ONE). Basically, it misses most anemic patients. Not great.

The bulbar conjunctiva - that clear membrane over the white of your eye - is the only microvascular bed you can observe without cutting through skin. It's like a tiny window into your circulatory system, except nobody brought the right magnifying glass until now.

Enter Video-to-Vessels (Yes, That's the Real Name)

Tamir Denis and colleagues at Tel Aviv University and Sheba Medical Center built a pipeline that takes high-magnification video of these conjunctival capillaries and compresses it into compact spatiotemporal representations - shrinking the data by roughly 200x while keeping the juicy hemodynamic details (Denis et al., 2026).

Here's the clever part: instead of treating the whole eye as one blob of pixels, they isolate individual vessels and feed each one into a modified ConvNeXt backbone. Think of it as giving every tiny blood vessel its own personal AI analyst. Then a cross-attention mechanism fuses vessel-specific thickness information (because vessel width matters when you're trying to estimate what's flowing through it), and the concatenated embeddings predict actual blood biomarker values.

On 224 participants with paired lab results, VesselNet hit a hemoglobin-based anemia ROC-AUC of 82.8% and a Spearman correlation of 0.47 for hemoglobin regression. It also managed a ρ of 0.46 for red blood cell count - meaning it's predicting two biomarkers from eye videos, not just sorting people into anemic/not-anemic buckets.

OK, 82.8% Doesn't Sound That High

Fair point. Other groups have gotten flashier numbers. A 2025 study using Vision Transformers on conjunctival images reported 98.47% accuracy for binary anemia classification (Scientific Reports, 2025). A smartphone-based approach using stacked CNNs hit an AUC of 0.97 (Healthcare Informatics Research, 2025). Even a multi-body-part network analyzing eyes, nails, and palms together reached 0.849 accuracy in retrospective testing (Journal of Imaging Informatics in Medicine, 2025).

But here's what makes Video-to-Vessels interesting: it's doing regression, not just classification. Telling someone "you're anemic" is useful. Telling someone "your hemoglobin is approximately 10.2 g/dL" is a different, harder, and arguably more useful task. And it's doing it from video - capturing how blood actually moves through vessels, not just what color they appear in a single frame.

The Ablation Study Is Where It Gets Spicy

When the team removed their local stabilization and segmentation-denoising steps, hemoglobin correlation dropped by 38% and RBC prediction dropped by 19%. Translation: the preprocessing pipeline that tracks and cleans up individual vessel segments isn't just nice-to-have - it's load-bearing. The AI needs clean, stabilized vessel footage to work, which makes sense when your input is a shaky close-up of someone's eyeball.

What This Actually Means for Your Next Doctor's Visit

Nobody's replacing your lab CBC with an eye video tomorrow. A cohort of 224 is promising but small, and a ρ of 0.47 means there's still plenty of noise. The current gold standard - a needle, a tube, and a Sysmex analyzer - remains untouchable for precision.

But consider the use cases where "good enough" beats "not available": remote clinics without phlebotomists, frequent monitoring for chemotherapy patients who are tired of being pincushions, or screening programs in regions where lab infrastructure is sparse. If you could point a specialized camera at someone's eye and get a reasonable hemoglobin estimate in minutes, that's not a replacement for the lab - it's a triage tool that didn't exist before.

The emerging field of "oculomics" - mining the eye for systemic health biomarkers - keeps finding that our eyes really do reveal more than we thought. Retinal scans predict cardiovascular risk. Conjunctival imaging detects anemia. Your eyeballs are basically tattling on the rest of your body.

For anyone working in computer vision for medical imaging, the vessel-level representation learning approach here is worth paying attention to. Processing tools like combb2.io already demonstrate how much detail AI can extract from visual data through enhancement techniques - applying similar image processing rigor to clinical video is a natural next step.

The Bottom Line

Video-to-Vessels proves that temporal hemodynamic information from conjunctival capillary videos adds real signal for blood biomarker estimation. The numbers aren't clinical-grade yet, but they're pointed in the right direction, and the architecture - per-vessel encoding with cross-attention fusion - is genuinely elegant. Your eyeball's tiny blood vessels have stories to tell. Turns out, you just need the right neural network to listen.

References

1. Denis, T., Sher, I., Praisman, E., et al. (2026). Towards noninvasive blood count using a deep learning pipeline from bulbar conjunctiva videos. NPJ Digital Medicine. DOI: 10.1038/s41746-026-02598-2

2. Non-invasive anemia detection from conjunctiva and sclera images using vision transformer. (2025). Scientific Reports. DOI: 10.1038/s41598-025-32343-w

3. Deep Learning Model-Based Detection of Anemia from Conjunctiva Images. (2025). Healthcare Informatics Research. PMID: 39973037

4. Khan, R., et al. (2025). Noninvasive Anemia Detection and Hemoglobin Estimation from Retinal Images. Translational Vision Science & Technology. PMID: 39847377

5. Real-time non-invasive hemoglobin prediction using deep learning-enabled smartphone imaging. (2024). BMC Medical Informatics and Decision Making. PMID: 38951831

6. BPANet: Deep Learning-Based Model for Non-invasive Hemoglobin Estimation via Body Parts Images. (2025). Journal of Imaging Informatics in Medicine. PMC: PMC11950610

7. Sheth, T.N., et al. Clinical examination for anemia detection: pallor assessment. PLOS ONE. DOI: 10.1371/journal.pone.0008545

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.