WorkDenTech
A clinical platform around a research AI
DenTech brings QCRI's oral-lesion detection model to clinics and research partners, in pilot. We lead the product and build the platform around it.
- Partner
- QCRI, which owns the AI model
- Sector
- Healthcare and research
- Our role
- Product and project leadership, and the platform around QCRI's model
- Status
- In pilot, for research use
The outcome
A pilot built to improve the model
DenTech is in pilot, for research use, connected to QCRI's model. The pilot is set up so every case does two jobs: it helps a clinic today, and it helps the model improve.
One workflow
Clinics submit cases and get the model's report.
Verified by experts
Diagnoses checked with blind second readings.
Consent on every case
Recorded, versioned and withdrawable.
Back to the model
De-identified data flows back to improve it.
The challenge
Real clinical data, with consent at every step
A promising AI model needs real clinical data to improve, and clinics need a tool that respects consent and privacy at every step.
The platform had to serve both, with clinical data and training data kept strictly apart.
What it had to do
- Real clinical data for the model
- Consent and privacy for clinics
- Clinical and training data kept apart
Our role
Product lead, and the platform around the model
We lead DenTech's product and project management, and we build the platform. We connected it to QCRI's model, and every report names the QCRI model as its source.
How the work ran
- 01
Product lead
DenTech's product and project management.
- 02
The platform
Cases, reports, second readings, consent and audit.
- 03
QCRI's model, connected
Each case goes to the model, and its report comes back to the clinic.
- 04
Tracked to release
Every change tracked from request to release, on a board QCRI can see.
- Now
In pilot
For research use, with clinics and research partners.
What users get
For clinics
- Upload photos of a lesion and get a one-page report from QCRI's model
- Blind second readings: a reviewer records their own read before seeing the first, so agreement is measured and never overwritten
- Disagreement between the model and the clinician recorded, not hidden
- A biopsy-pending hold that keeps a case out of the training data for up to 30 days, until the result is in
- Consent attested by each clinician, versioned and withdrawable at any time
01
Clinic data
Each clinic and partner sees only its own cases.
02
Training data
Kept apart, to improve the model.
For research partners
- Bulk uploads, with every image checked on its own, failures listed with reasons and duplicates flagged
- Training data de-identified and kept apart from clinic data
- One platform for many clinics and partners, each seeing only its own cases, with an audit trail it can review
Where it stands
In pilot, for research use
Clinics and research partners take part in the pilot, each seeing only its own cases.
- In pilot
- for research use
- Partners
- clinics and research institutions