Source study found

Story checked

New AI system accelerates rabies virus detection in cell cultures (opens in a new tab)

news-medical.net · 2026-10-12

Short answerEvidenceSource

Short answer

Mostly not supported

Mostly not supported.

2 claims go further than the study. One other point was not covered by the paper.

1 supported
2 overstated
1 not covered

Checked against the study summary. The full text wasn't available, so some details couldn't be settled either way.

Share this check

Follow the evidence trail

The story

New AI system accelerates rabies virus detection in cell cultures

news-medical.net · 2026-10-12

The story’s checkable claims.

Read the original story

NewsLink checks it

Mostly not supported

Two of four claims overstate the study. One of four checks out. One claim the study doesn't address.

1 supported
2 overstated
1 not covered

Open claim evidence

The source study

Deep learning enables rabies virus detection and antibody quantification in cell culture

Scientific Reports · 2026

Open source record Open in Tessa

Source paper

Source layer

The 2 papers the story cites

Source study separated from background citations.

The research anchor for the report.

The study this story reportspresented as the new finding
Deep learning enables rabies virus detection and antibody quantification in cell culture
Scientific Reports · 2026
PubMed (opens in a new tab)Full Tessa analysis (opens in a new tab)
The study this story reportspresented as the new finding
Deep learning enables rabies virus detection and antibody quantification in cell culture
Scientific Reports · 2026
PubMed (opens in a new tab)Full Tessa analysis (opens in a new tab)

Then inspect each claim

Evidence layer

Claim by claim

Each claim gets a verdict. Expand it to see the evidence directly below.

Reading mode

Scan verdicts. Open evidence only when needed.

Browse by verdict

4 claims in this story

Showing all 4 claimsChoose a verdict to focus the list.

Claim 1 of 4OverstatedResearchers at the University of Surrey and the APHA developed a world-first AI system to accelerate rabies virus detection in cell cultures and help improve surveillance of rabies.View evidenceHide evidence

As statedworld-first

Why this verdict

The abstract-level profile supports a 'first AI-based workflow' for detecting rabies virus infection in BHK-21 cell cultures, so the world-first/first-of-kind detection element is supported. However, the headline framing that it will 'accelerate' detection and improve surveillance goes beyond the abstract evidence, which reports model development and performance but not measured speed gains, surveillance outcomes, or deployment impact. The University of Surrey/APHA attribution is not verifiable from the supplied profile.

Study evidence

MambaOut achieved very high well-level diagnostic performance in detecting rabies infection from BHK-21 fluorescence images.well-level accuracy 0.996; F1-score 0.993; AUC-ROC 0.999

“the first AI-based workflow for detecting rabies virus infection in BHK-21 cell cultures”

Claim 2 of 4OverstatedThe study says the AI workflow can identify rabies virus infection in cell cultures and accurately measure antibody concentration levels, replacing manual microscope review with a faster, more consistent process.View evidenceHide evidence

Why this verdict

The paper profile supports that the workflow detects infection in cell-culture fluorescence images and is used to quantify antibody titres/concentrations, with high well-level infection-detection performance and an abstract claim of complete agreement for antibody concentrations. But the story's stronger claim that this replaces manual microscope review with a faster, more consistent process is not demonstrated in the abstract-level evidence; speed, replacement of specialists, and operational consistency were not quantified.

Study evidence

MambaOut achieved very high well-level diagnostic performance in detecting rabies infection from BHK-21 fluorescence images.well-level accuracy 0.996; F1-score 0.993; AUC-ROC 0.999

“the first AI-based workflow for detecting rabies virus infection in BHK-21 cell cultures”

Study evidence

MambaOut predictions of antibody concentrations on unseen images from the WHO/WOAH fluorescent antibody virus neutralisation assay showed "complete agreement with ground truth" as reported in the abstract.Complete agreement (as stated in abstract)

“detecting rabies virus infection in BHK-21 cell cultures and quantifying antibody titres”

Claim 3 of 4Not coveredThe article says the technology could help laboratories analyze samples more quickly and consistently in rabies-endemic countries, could measure antibody levels used to assess vaccine protection and pet-travel requirements, and may later be developed to detect virus in infected brain tissue or be adapted to other infectious diseases.View evidenceHide evidence

Why this verdict

The abstract-level profile supports practical relevance for antibody concentration/titre prediction and reports testing on unseen FAVN assay images. However, the broader story claims about use in rabies-endemic countries, faster and more consistent laboratory analysis, later development for infected brain tissue, and adaptation to other infectious diseases are not available in the supplied abstract-level profile. Because these are future/deployment claims that may appear outside the abstract, they are not verifiable at this evidence depth.

Study evidence

MambaOut achieved very high well-level diagnostic performance in detecting rabies infection from BHK-21 fluorescence images.well-level accuracy 0.996; F1-score 0.993; AUC-ROC 0.999

“the first AI-based workflow for detecting rabies virus infection in BHK-21 cell cultures”

Study evidence

“detecting rabies virus infection in BHK-21 cell cultures and quantifying antibody titres”

Claim 4 of 4SupportedUsing more than 2,300 fluorescence images with over 36,000 annotated image sections, the team trained deep learning models; the best-performing model correctly classified 99.6% of samples and almost perfectly distinguished infected from uninfected cells.View evidenceHide evidence

As stated99.6%

Why this verdict

The profile states that ResNet, Swin Transformer V2, and MambaOut were trained on 2,344 fluorescence images and 36,822 annotated patches, and that MambaOut achieved 0.996 well-level accuracy, 0.993 F1-score, and 0.999 AUC-ROC. The story's 'more than 2,300' and 'over 36,000' figures and 99.6% best-model classification claim match the abstract-level evidence, though the metric is specifically well-level accuracy rather than a generic sample-level result.

Study evidence

MambaOut achieved very high well-level diagnostic performance in detecting rabies infection from BHK-21 fluorescence images.well-level accuracy 0.996; F1-score 0.993; AUC-ROC 0.999

“the first AI-based workflow for detecting rabies virus infection in BHK-21 cell cultures”

Then look for missing context

Context layer

What the story left out

Important study details the story did not include.

The antibody-concentration applicability test used unseen images from the WHO/WOAH-recommended FAVN assay, with MambaOut reported as showing complete agreement with ground truth.
The story reflects the broad antibody-measurement use case but does not convey the specific FAVN assay context, the unseen-image applicability test, or the exact 'complete agreement' abstract claim.
From External validation / transportability test (FAVN)
The abstract does not provide key evaluation details such as data splits, held-out/test sample sizes, cross-validation design, confidence intervals, or per-class error analyses.
The story mentions reliance on manual microscope review but does not note these model-evaluation limitations, which matter for interpreting the reported 99.6% performance.
From Deep learning image-classification training and multi-level evaluation
The abstract does not establish generalisability to other laboratories, microscopes, staining protocols, viral strains, or broad endemic-country deployment settings.
This limitation is not mentioned, while the story makes broader accessibility/endemic-setting claims that are not supported at abstract depth.
From Deep learning image-classification training and multi-level evaluation
For antibody quantification, the abstract gives no sample count, numerical agreement metric, uncertainty estimate, or definition of 'complete agreement'.
The story reports accurate antibody measurement but does not include the abstract-level limitations that make the strength of that agreement difficult to assess.
From Deep learning inference on fluorescence microscopy images for antibody titre/concentration prediction; External validati

4 things the story did carry across

The paper presents the work as the first AI-based workflow for detecting rabies virus infection in BHK-21 cell-culture fluorescence images.
The main model-development evidence used 2,344 fluorescence images and 36,822 annotated patches, comparing ResNet, Swin Transformer V2, and MambaOut with patch-, image-, and well-level evaluation.
MambaOut achieved the headline infection-detection performance: 0.996 well-level accuracy, 0.993 F1-score, and 0.999 AUC-ROC.
The workflow was also used for rabies-neutralizing antibody titre/concentration prediction from fluorescence assay images.

Then read the study layer

Study layer

Study at a glance

Scan the study first. Expand only the parts you want to inspect.

Pieces of work

Evidence read

study summary

Lead result

in silico

1Lead resultin silicoDevelop and train deep-learning models to detect rabies virus infection from fluorescence microscopy images of BHK-21 cell cultures, enabling automated well-level infection calls.Deep learning image-classification training and multi-level evaluationExpand

In plain English

Developed and internally evaluated deep-learning classifiers (ResNet, Swin Transformer V2, and MambaOut) to detect rabies virus infection in fluorescence microscopy images of BHK-21 cell cultures. Models were trained on a labeled dataset of 2,344 images (36,822 annotated patches) and assessed at patch, image, and well aggregation levels; MambaOut achieved the highest well-level performance and was further tested on unseen images from a WHO/WOAH-recommended fluorescent antibody virus neutralisation (FAVN) assay for antibody concentration prediction.

Key findings

MambaOut achieved very high well-level diagnostic performance in detecting rabies infection from BHK-21 fluorescence images.well-level accuracy 0.996; F1-score 0.993; AUC-ROC 0.999
Three architectures (ResNet, Swin Transformer V2, MambaOut) were trained and compared on a large annotated dataset, with evaluation at patch, image, and well aggregation levels.

“the first AI-based workflow for detecting rabies virus infection in BHK-21 cell cultures”

What this piece can’t prove

3 further details could not be confirmed from the summary.

2in silicoUse the same AI workflow to quantify rabies-neutralizing antibody titres/concentrations from cell-culture assay images (antibody quantification use case).Deep learning inference on fluorescence microscopy images for antibody titre/concentration predictionExpand

In plain English

The authors applied the trained deep-learning workflow to predict rabies-neutralizing antibody concentrations/titres from fluorescent cell-culture assay images (WHO/WOAH-recommended fluorescent antibody virus neutralisation assay). The MambaOut model produced predictions that the abstract reports as showing complete agreement with ground-truth antibody concentrations on unseen assay images.

Key findings

MambaOut predictions of antibody concentrations on unseen images from the WHO/WOAH fluorescent antibody virus neutralisation assay showed "complete agreement with ground truth" as reported in the abstract.Complete agreement (as stated in abstract)

“detecting rabies virus infection in BHK-21 cell cultures and quantifying antibody titres”

What this piece can’t prove

Unclear whether the evaluation was external/independent or limited to held-out images from the same dataset/laboratory.

2 further details could not be confirmed from the summary.

3in silicoExternally test applicability/generalizability on unseen images from the WHO/WOAH-recommended fluorescent antibody virus neutralisation (FAVN) assay and assess agreement with ground truth antibody concentrations.External validation / transportability test (FAVN)Expand

In plain English

The study performed an external/held-out applicability test by running trained deep-learning models on unseen fluorescence images from the WHO/WOAH-recommended fluorescent antibody virus neutralisation (FAVN) assay to predict antibody concentrations; MambaOut showed complete agreement with ground-truth antibody concentrations in this external test as reported in the abstract.

Key findings

On unseen FAVN assay images, MambaOut produced predictions that the authors report as in complete agreement with ground-truth antibody concentrations.complete agreement

“To assess applicability, models were tested on unseen images from the WHO/WOAH-recommended fluorescent antibody virus neutralisation assay.”

What this piece can’t prove

Abstract-only reporting: no numeric details on external test sample size or distribution are provided.
The abstract does not specify the metric or criteria used to define 'complete agreement' between predicted and ground-truth antibody concentrations.
Provenance and collection details for the FAVN test images (e.g., different labs, operators, or instruments) are not described, limiting assessment of transportability.

1 further detail could not be confirmed from the summary.

Finally, the search trail

Method layer