← Back to Home

Teaching AI to Read Rapid Tests with Synthetic Images: SynSight Demonstrates Performance on HIV and COVID-19 Tests

Training a model to interpret rapid tests may not require collecting a vast library of real photographs first. SynSight performed well in HIV and COVID-19 rapid test validation after training on simulated images, but accurate image interpretation remains some distance from reliable clinical diagnosis.

By SURL BioNews

Rapid tests take testing out of the laboratory, but someone still needs to interpret the results. To have artificial intelligence assist with interpretation, developers often first need to collect large numbers of photographs and label them individually—a particularly significant hurdle for newly launched tests or those with limited image data. A new study tries a different starting point: generating training images by computer, then testing whether the model can interpret real rapid tests.

A study published on November 10 in *npj Digital Medicine* introduces SynSight, a workflow for rapid test image segmentation and classification trained on synthetic data. The journal abstract reports sensitivity of 98% and specificity of 99% in HIV rapid test validation, with accuracy of up to 99% for COVID-19 rapid tests. The version currently available is an advance release that has been accepted following peer review and awaits further editing.

An early abstract from University College London’s (UCL) 2025 research event provides a clearer outline of the method’s technical approach. Elliott Rogers and colleagues trained models on mechanistically generated synthetic images, linking real-time object detection with result classification: first locating the test region in a photograph that needs analysis, then interpreting the result. The research aims to reduce model training’s reliance on libraries of real images.

The early UCL abstract lists HIV sensitivity and specificity of 98.2% and 99.2%, respectively. The COVID-19 evaluation covered two rapid antigen tests and one rapid antibody test, with accuracy of 99%, 97%, and 95%, respectively. These are earlier-stage data and cannot be treated directly as the complete detailed results of the journal version. They also remind readers that “up to 99%” does not mean every rapid test achieved the same performance.

The appeal of synthetic data is that developers may not need to build a new, large, expensive training photograph library involving sensitive health information for every new test. However, dispensing with real photographs as training material does not mean real-world validation can be omitted. Whether simulated images can capture variations in actual photography and test performance remains key to how far this approach can go.

When interpreting the figures, it is also necessary to distinguish between “interpreting a rapid test image” and “determining whether a person is infected.” An image model’s interpretation performance cannot be taken directly as the diagnostic ability of the entire testing process to identify infected people. The available journal abstract does not specify the validation sample size, confidence intervals, or complete performance under different conditions of use. It therefore does not yet provide enough evidence to infer that the model is equally reliable across different brands, phones, or home environments.

SynSight offers a feasible direction for lowering development barriers. Before it can enter routine testing services, the next steps still require clarifying which tests and photographic conditions it supports, how to handle results that cannot be interpreted reliably, and the applicable medical device regulatory requirements. The study data support further validation of this approach, but they do not yet establish that it can independently perform diagnostic work.

References

  1. npj Digital Medicine
  2. University College London