Read test results from images instead of reading them one by one
- 100%Processed on your own site
If any of this sounds familiar
- Reading test results one by one takes a long time and becomes a bottleneck when volume is high
- Different people read results to different standards, so results come out inconsistent
- You want a system that learns your organization's own expert standard, not a generic one that doesn't fit your context
- Test-result data is often business-confidential, and you don't want images leaving the organization to be processed
How the system works
Photograph test results in a consistent, standardized way
Set up the camera and lighting so every test result is photographed the same way each time — this makes the model easier to train and more accurate.
Have an expert label the initial examples
The person most knowledgeable about this at your organization decides what each image means. The system learns your organization's own standard, not a generic one.
Train a model to classify the result from the image
The labeled images are used to train an image classification model. The more examples that cover real variation, the more accurate it gets.
Use it and keep improving
Once it's in real use, new images keep improving the model's accuracy going forward.
The signature visual for this page has not been built yet (CT-031/032)
Technical specifications
- Model
- Vision classification / object detection (local)
- Where it runs
- On your organization's own hardware only
Hardware used
Hardware package pricing is still being worked out — it depends on how many result types and how many images need training. Get in touch for an estimate based on your actual needs.
See full specs and prices on the hardware pageConnects to your existing records system
Readings are sent to your existing QC/LIMS records system via API or file import, so nothing has to be re-keyed by hand.
Frequently asked questions
What this system cannot do
- It always needs an expert to label the initial examples before a model can be trained — it doesn't work on day one with no data.
- A result pattern it has never seen before can be misclassified — it needs enough example coverage before real use.
- Still in development, with no customer or pilot running it yet — there are no published accuracy or speed figures.