Collaboration · Medical technology
Sascan Meditech
We’re building the AI layer for a portable device that aims to find cervical precancer and treat it in the same visit. Here is exactly where that work stands, and what is not yet true.
Challenge
Cervical precancer screening in rural and remote Australia is a relay of appointments: screen, wait, diagnose, wait, treat. Every wait is a chance to lose the patient, and the communities with the least access are the ones losing the most.
Sascan Meditech set out to collapse that relay into a single visit with a portable, smartphone-connected “see and treat” device. That design has one hard dependency: somebody has to read the fluorescence imagery in the room, in real time, without a specialist present. That is the problem we were brought in for.
Solution
ALTDATA supplies the AI analytics layer, and only that. The split matters, so here it is plainly:
- Ours. Automated classification of fluorescence images into lesion grade and risk category, referenced against expert-annotated data, plus tracking of fluorescence intensity during treatment to help inform when a session has done its work.
- Sascan’s. The imaging hardware and probe, the photodynamic therapy light-delivery system, the 5-ALA photosensitiser, the clinical trial programme, and the TGA regulatory pathway for the combined device.
Sascan Meditech is not a first-time builder. Its oral-cancer screening device, OralScan, is already on the market; CerviScan is in Phase-3 clinical validation; and the company holds ISO 13485:2016 certification. We are the analytics component inside someone else’s regulated device, a role we think more AI companies should be honest about wanting.
Technology
The physics is the appeal. Under 405 nm violet excitation, protoporphyrin IX (which accumulates in precancerous cells and is amplified by a topically applied 5-ALA photosensitiser) fluoresces at longer wavelengths, with peaks around 635 nm and 705 nm. Precancerous tissue effectively lights itself up.
That gives a classifier a real physical signal to work with rather than a texture it has to guess at. ALTDATA’s component is designed to grade that signal, CIN 1–3 staging for cervical tissue, in real time on a device, correlated against histopathology and cytology ground truth.
Where it stands
Concept stage. No model has been built, trained or benchmarked yet, so there is nothing here to report as a result, and we won’t invent one.
What is real: an active collaboration with a named device partner that already has products in market and in trials. What comes before any model does: a data-sharing agreement for access to expert-annotated fluorescence imagery, and clarity on how an AI component is regulatory-classified inside Sascan’s TGA pathway for a combined diagnose-and-treat device.
When there are sensitivity and specificity numbers, they will appear here with their measurement basis attached. Until then, this section stays empty of figures on purpose.
Why we’re in it
This is the ALTDATA thesis in someone else’s hardware. The bottleneck is not model architecture. It is that expert-graded fluorescence imagery is scarce, expensive and slow to produce. That is the same data gap our other work attacks, in a domain where closing it has an obvious human dividend.
It also puts us where we want to be: supplying a defensible analytics component into a regulated device, with a partner who already knows what regulators ask for.
What’s next
Formalise data access, build a first lesion-classification model, and benchmark it honestly against real annotated imagery. We’ll publish what it does, including where it fails.