Researchers at Tel Aviv University and Mashiva have developed an artificial intelligence system that analyzes short videos of blood vessels in the conjunctiva and estimates hemoglobin levels and red blood cell counts. In a trial of 224 participants, it detected anemia with an AUC of 82.8%—a promising result, but still far from a replacement for a blood test.
A blood test is one of the most common medical tests, but it is also one of the most basic medical procedures that still requires a puncture, a sample, and laboratory equipment. Researchers from Tel Aviv University and Sheba Medical Center are now proposing a different approach: looking at blood while it is still flowing through the body — through the tiny blood vessels visible on the surface of the eye.
The team developed a system of computer vision וartificial intelligence which analyzes short videos of blood vessels inלחמית the eye and tries to predict from them the concentration ofhemoglobin and the number of red blood cells.
The study was published in the journal npj Digital Medicine. and was conducted under the leadership of Tamir Dennis, in collaboration with the research groups of Prof. Chaim Suchovsky from the School of Physics and Astronomy and Prof. Lior Wolf from the School of Computer Science and Artificial Intelligence at Tel Aviv University, together with researchers from Sheba Medical Center, including Dr. Yifat Sher and Prof. Yigal Rothenstreich. ([Nature][1])
The blood vessels in the eye are a window to the circulatory system.
The conjunctiva is the transparent membrane that covers the white part of the eye. Underneath it runs a dense network of blood vessel Very small, relatively easy to photograph from the outside.
Unlike the skin, which hides some of the optical information coming from the blood vessels beneath it, in the eye, blood vessels and even the movement of blood cells through them can be observed relatively directly.
The researchers took advantage of this to develop a system they call Video-to-VesselsThe system receives a high-magnification video of the conjunctiva, stabilizes it, identifies the blood vessels, and converts each blood vessel into a compact representation that describes how the optical information and flow change over time.
The process reduces the amount of information in the video by about 200 times, while attempting to preserve important hemodynamic information. ([Nature][1])
Ten seconds from each eye
The study involved 224 women and men. Each participant was filmed making approximately ten-second videos of the twoEyes, and a regular blood test was also taken on the same day.
The laboratory test served as the "benchmark" against which the system's predictions were compared.
After processing the videos, the blood vessel representations were fed into a deep learning model called VesselNetThe model does not look at a single blood vessel, but combines information from several blood vessels in the same person—along with the thickness of the blood vessels—to predict hemoglobin concentration or red blood cell count. ([Nature][1])
How good was the forecast?
For hemoglobin concentration, a Spearman correlation of 0.47 between the predictions and the laboratory tests. For the red blood cell count, a correlation of 0.46.
These are moderate correlations — meaning the system is not yet able to replace an accurate quantitative blood test.
However, when hemoglobin prediction was used to distinguish between people with anemia and people without anemia, was obtained ROC-AUC of 82.8%This is a measure of the discrimination ability of a classifier: a value of 50% is essentially equivalent to a random guess, and 100% reflects perfect separation.
Therefore, it is important to be precise: the result does not mean that the system “measured hemoglobin with 82.8% accuracy” or that 82.8% of all tests were correct. It is the AUC of the screening test for anemia detection. ([Nature][1])
The thin blood vessels provided more information.
The researchers also looked at which blood vessels contributed most to the prognosis.
Surprisingly, removing the thinner blood vessels impaired hemoglobin prediction performance. The researchers suggest that small blood vessels—tiny capillaries and venules—may provide a particularly rich optical signal because blood cells often travel in narrower lanes, allowing for better detection of changes in flow and hemoglobin levels.
The finding does not mean that you can simply "count blood cells" with a camera. The model analyzes complex spatial and temporal patterns originating from blood vessels and the flow through them.
Why do blood tests need to be replaced at all?
blood tests Conventional ones are accurate, relatively inexpensive, and very common in countries with developed health systems. Therefore, the new technology currently offers no reason to give them up.
The potential benefit is found primarily in places where blood sampling is more complex: remote areas, clinics without a laboratory, frequent patient monitoring, extensive screening tests, or people who need repeated measurements.
Anemia is a particularly pertinent example. Low hemoglobin levels can reduce the amount of oxygen reaching tissues, and the problem is very common worldwide. A system that allows for rapid, noninvasive screening could help identify people who then need a full medical examination. Tel Aviv University states that the future vision is to develop a small, accessible device that could be used in clinics and perhaps at home. ([Technion][2])
Not yet "blood count through the camera"
The title of the scientific article talks about a step towards non-invasive blood counts, but the current study only examined two indicators: Hemoglobin concentration and number Red blood cells.
A complete blood count includes many other measures, including white blood cells, platelets, hematocrit, and various blood cell characteristics. The study does not show that all of these can already be measured from an eye exam.
Considerable progress is still needed for both of the indicators tested. The correlation with laboratory tests is not high enough to replace clinical measurement, and the system will need to be tested in large and diverse populations and under different medical conditions.
The researchers themselves define the work as a first step towards a non-invasive system, rather than a ready-made medical product. ([Nature][1])
Combination of optics and artificial intelligence
The work demonstrates a broader trend in medicine: using relatively simple sensors to extract complex physiological information through image processing and machine learning.
Instead of developing a sensor that directly measures hemoglobin concentration, the researchers let the model learn which characteristics within hundreds of imaged blood vessels are related to the laboratory values.
Such an approach could turn a camera into a kind of medical sensor—but it also requires caution. A model trained on a particular population will not necessarily maintain the same level of performance across ages, diseases, medications, imaging conditions, or other populations.
Therefore, the path from a scientific article to a device that can be relied upon in medical decisions is still long.
Questions and Answers
Is it already possible to perform a blood test using a photo of the eye? No. The study demonstrates research feasibility for predicting only two indicators, and it is not yet a substitute for a medical blood test.
What was the system able to measure? She estimated hemoglobin concentration and red blood cell count using conjunctival blood vessel films.
What does an AUC of 82.8% mean? This is a measure of the system's ability to distinguish between people with anemia and people without anemia. It does not mean that the system measures hemoglobin with 82.8% accuracy.
Why photograph the eye? Because the blood vessels in the conjunctiva are close to the surface, they and the blood flow within them can be photographed relatively directly and without a puncture.
More on the subject on the science website
- Orsens launches a system for continuous and non-invasive measurement of hemoglobin level and blood oxygen saturation
- Can an algorithm make decisions without a doctor? Israel to test autonomous medical artificial intelligence
- When the machine wears a lab coat: Artificial intelligence enters the lab
- A drop of blood to diagnose cancer
- Ultrasound interpretation technology
The scientific article
Tamir Denis et al., Towards noninvasive blood count using a deep learning pipeline from bulbar conjunctiva videos, npj Digital Medicine 9, 431 (2026). The full article DOI
Link to image: https://www.tau.ac.il/sites/default/files/media_server/General/DK/Eye-Vein-Scan-580.jpg
Image caption: The process of analyzing blood vessels in the conjunctiva: A short video of the eye is processed for information about blood vessels and blood flow, from which the artificial intelligence system attempts to estimate blood indices. Credit: Tel Aviv University.
OLD: Artificial intelligence system analyzes blood vessels in the conjunctiva to assess hemoglobin and red blood cells
Image file name: `eye-vessels-noninvasive-blood-test-ai.jpg`
[1]: https://www.nature.com/articles/s41746-026-02598-2?utm_source=chatgpt.com "Towards noninvasive blood count using a deep learning pipeline from bulbar conjunctiva videos | npj Digital Medicine" [2]: https://www.tau.ac.il/research/Eye-Vein-Scan-26 "Looking into the Blood with Your Eyes: Towards Needle-Free Blood Testing | Tel Aviv University | Tel Aviv University"
For the original publication: [Tel Aviv University – Looking at Blood with the Eyes: Towards Needle-Free Blood Testing]
