Case study · Healthcare
Cancer detection and medical imaging
Inside AI Research Labz, where the detection models are benchmarked before they are claimed.
- Team
- AI Research Labz
- Focus
- Detection, classification, prediction
- Sector
- Healthcare research
- Status
- Ongoing, work in progress
Benchmarks
Where the research stands
On the full benchmark lung cancer dataset.
Skin, ovarian, kidney, brain, cervical, lung, colon and prostate.
A 3D segmentation model and automated detection workflows are active work, not shipped claims.
The research targets existing clinical data rather than a standalone tool.
The brief
Research that states its benchmark
Detection accuracy is the easiest number in medical AI to quote and the easiest to quote misleadingly. A figure means something only alongside the dataset it came from, and a model that performs on one cancer type tells you very little about the next.
So this page is deliberately narrow about what is established and what is not. 97% accuracy on the full benchmark lung cancer dataset is a measured result. The wider programme — eight cancer types under benchmark, a 3D segmentation model, automated workflows — is in progress.
A number without its dataset is marketing. The dataset is the claim.
In progress
What is being built now
Two things are active. A 3D segmentation model, which matters because clinical imaging is volumetric and a per-slice model discards the dimension clinicians actually read. And automated workflows for cancer detection, which is the difference between a model and something a department could use.
Both are being built toward integration paths into existing EHR and MRI data. That constraint is doing real work: it rules out approaches that only function on clean research datasets, which is where a lot of published medical AI quietly lives.
AI Research Labz is Wow Labz's applied AI team, and this is a live programme rather than a delivered project. Treat the in-progress items as in progress.
Research threads
Three lines of work
Detection and benchmarking
Detection benchmarked across skin, ovarian, kidney, brain, cervical, lung, colon and prostate cancers. Breadth is the point: a model is only as trustworthy as the range it has been tested over.
Classification with neuro symbolic networks
Combining learned representations with symbolic structure, aimed at classification that can be reasoned about rather than only measured.
Prediction from gene expression
Working upstream of imaging entirely, predicting from gene expression data instead of from a scan.
Stack
What it was built with
Questions
The research, answered
What accuracy has been achieved?
97% accuracy on the full benchmark lung cancer dataset. That figure is specific to that dataset and that cancer type; the wider benchmarking programme across other cancers is ongoing.
Which cancers are being benchmarked?
Skin, ovarian, kidney, brain, cervical, lung, colon and prostate.
What is still in progress?
A 3D segmentation model and automated workflows for cancer detection, both being built toward integration paths into existing EHR and MRI data. These are active work rather than delivered capabilities.
Is this a product?
No. It is a research programme inside AI Research Labz, Wow Labz's applied AI team. The benchmarked result is real; the surrounding workflow and integration work is ongoing.
Let's talk
Got a developer product nobody finishes integrating?
Tell us what you are building. We will tell you how we would ship it.
