AI reads rabies test cultures with 99.6% well-level accuracy

Reading fluorescence images is a labour-intensive part of laboratory rabies testing. Researchers at the Animal and Plant Health Agency and the University of Surrey have tested whether deep learning can take on that visual task, reporting 99.6% accuracy for classifying culture wells.[1]

The peer-reviewed Scientific Reports study, published in June and highlighted again by Surrey in October, used 2,344 fluorescence images divided into 36,822 annotated patches. The material came from BHK-21 cell cultures, a laboratory setting used in rabies work. The researchers compared three network architectures—ResNet, Swin Transformer V2 and MambaOut—at patch, image and culture-well levels. MambaOut was the best-performing system across the reported comparisons.[1][2]

From patches to a culture well

At the culture-well level, MambaOut achieved 0.996 accuracy, an F1 score of 0.993 and an area under the receiver-operating-characteristic curve of 0.999. Those are results on an evaluated image-classification task. They are not percentages of people with rabies correctly diagnosed, nor do they state how many infections were caught before symptoms developed. Changing the denominator from a culture well to a human patient would change the meaning of the statistic.[1]

The team went beyond a simple infected-versus-uninfected exercise. It also tested unseen images from a fluorescent antibody virus neutralisation assay of the type recommended by international health and animal-health organisations. In the reported evaluation, the model’s predicted antibody concentrations agreed with the reference results. That matters because antibody measurement is relevant to vaccine assessment and some animal travel requirements, where a repeatable laboratory workflow has real operational value.[1]

An operational opportunity for laboratories

The useful prospect is a more consistent reading process when laboratories handle many images. Automating this stage could free skilled staff for preparation, ambiguous results and quality control. Whether it shortens a full assay depends on the rest of the laboratory workflow as well as the classifier.[1]

The published abstract establishes the dataset size and central results, but does not detail a multisite routine-service evaluation. The finding should therefore be understood as evidence for a specific laboratory automation task. It does not demonstrate a reduction in human rabies deaths or replace the other steps needed to prevent infection.[1]

The extension to antibody measurement makes the work particularly useful to follow: the same visual workflow could support both detecting infection in cultures and quantifying a vaccine-related response. Reproducing those results under routine laboratory conditions is the step that would turn an impressive classifier into dependable capacity.[1]

Report an error or suggest a correction