ScienceNot yet confirmed elsewhere1 publisher2 min readPublished
Surrey and APHA's best deep learning model classified 99.6% of rabies cell-culture samples correctly
Surrey and APHA researchers trained a model that classified 99.6% of rabies cell-culture samples correctly, according to a Scientific Reports study. The authors present it as support for hand-read microscopy, with a brain-tissue version still in development.
The Scientist · Science desk
What happened
- The same workflow also measures antibody concentrations, and the authors describe it as the first AI-based workflow to do both jobs in cell cultures.
- The team trained several deep learning models on more than 2,300 fluorescence images that held over 36,000 annotated image sections.
- Current diagnosis and antibody measurement depend on specialists examining thousands of microscope images by hand, which is slow, labor-intensive and open to differences in interpretation.
- Rabies causes around 59,000 human deaths a year even though vaccination can almost entirely prevent it.
Why it matters
- capability If the accuracy holds on other samples, the same model reads every image the same way, and reader-to-reader differences in interpretation drop out of the process.
- decision A laboratory weighing the tool has to decide whether to run it beside its own specialists on local samples first, because the published figure comes from one team's study.
- precedent Professor Miroslaw Bober says the framework could extend to other infectious diseases that depend on similar laboratory tests, so the rabies result also tests whether the approach generalises.
The 99.6% figure needs a denominator. The 2,300-plus training images average roughly 16 annotated sections each [6], yet the annotation is described at section level [3] while the accuracy is reported for samples [4]. Those may be different units. At 99.6%, about 4 calls in every 1,000 are wrong [5].
The account does not say how the images were divided between training and testing, whether any came from outside the originating laboratories, or what figure backs the claim that antibody concentrations are measured accurately [7]. So it cannot show how the model fares on someone else's samples. Antibody levels are what vaccine-protection checks and pet-travel requirements rely on [8], and the 99.6% describes infection classification [4].
The endemic-country case is stated as a possibility: the technology could help laboratories where specialist expertise is limited [8]. Joan Manuel Amaya Cuesta, a viral zoonoses research scientist, said the workflow offers "a faster and more consistent way to analyze samples while reducing the reliance on manual interpretation" [9].
Human readers differ in interpretation [2], and the model was scored against annotated sections [3], so part of the 0.4% could reflect disagreement in the labels, not model error.
The hope for diagnostic laboratories is tied to a later version. "This technology is now being further developed to detect the virus in infected brain tissue, which we hope will transform international rabies diagnostic laboratories," said Dr. Lorraine M. McElhinney, who leads APHA's National Reference Laboratory for Human and Animal Rabies [10].
What to watch
- A breakdown of the test set: how images were split, and whether any came from laboratories outside the originating ones.
- A published agreement figure for the antibody measurements against specialists' readings.
- APHA results from the brain-tissue version McElhinney describes as under development.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence50
- Adoption
- Insufficient
- Hype gap+30
- Incentives
- Insufficient
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Despite being almost entirely preventable through vaccination, rabies causes around 59,000 human deaths every year.
- [2]
Current practices for diagnosis and measuring protective antibodies after vaccination rely on specialists manually examining thousands of microscope images, which can be slow, labor-intensive and open to differences in interpretation.
- [3]
The team trained several deep learning models using more than 2,300 fluorescence images containing over 36,000 annotated image sections.
- [4]
The best-performing model correctly classified 99.6% of samples and almost perfectly distinguished infected from uninfected cells, avoiding false alarms.
- [5]
A 99.6% correct classification rate leaves 0.4% incorrect, about 4 in every 1,000.
- [6]
Over 36,000 annotated sections across more than 2,300 images averages roughly 16 sections per image.
- [7]
In a Scientific Reports study, the team created what it describes as the first AI-based workflow capable of identifying rabies virus infection in cell cultures and accurately measuring antibody concentrations.
- [8]
The technology could help laboratories analyze samples more quickly and consistently, particularly in countries where rabies remains endemic and specialist expertise is limited; it could also reliably measure antibody levels used to assess vaccine protection and support international pet travel requirements.
- [9]
"Our AI workflow provides laboratories with a faster and more consistent way to analyze samples while reducing the reliance on manual interpretation," said Joan Manuel Amaya Cuesta, viral zoonoses research scientist.
ReportedInsufficientSource: Joan Manuel Amaya Cuesta, viral zoonoses research scientist2 sources— create a free account to open themView cited source - [10]
"This technology is now being further developed to detect the virus in infected brain tissue, which we hope will transform international rabies diagnostic laboratories," said Dr. Lorraine M. McElhinney, head of the National Reference Laboratory for Human and Animal Rabies at the APHA.
- [11]
Professor Miroslaw Bober said that while the approach was demonstrated for rabies, the same framework could be applied to other infectious diseases that rely on similar laboratory tests.
ReportedInsufficientSource: Professor Miroslaw Bober, professor of video processing and machine intelligenceView cited source
Sources
1 independent publisher whose own reporting we read for this story.
- phys.orgAI system could transform rabies testing around the world
1 article · October 12, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Entities
- University of SurreyFollow
- Animal and Plant Health AgencyFollow
- Scientific ReportsFollow
- Miroslaw BoberFollow
- Lorraine M. McElhinneyFollow
- Phys.orgFollow