Upload a CT
Drop in an abdominal CT series. The model needs the CT alone. Radiology reports help only during training, so none is needed at the point of use.
Research preview · Built in Sydney
Tumour edges, drawn the way they actually are.
We're building toward a viewer for radiologists: upload an abdominal CT and see the pancreatic tumour highlighted with a soft, shape-faithful contour, and an honest band where the edge is uncertain. We're in research preview, and we publish our evidence as we go.
Research preview · not for clinical use
AXIAL
ABDOMEN WINDOW
ABDOMEN WINDOW
SYNTHETIC RENDER
NOT PATIENT DATA
NOT PATIENT DATA
Soft contour · graded band
CT-only inference
Concept interface. The slice is procedurally drawn and the contour is illustrative. Neither is model output.
Pancreatic CTCT-only inferenceSoft boundariesEvidence-firstBuilt in Sydney
The product we're building
From scan to contour, in three steps.
The workflow we're designing toward. Each step maps to a piece of research that is under way; none of it is a finished clinical product.
Delineate the tumour
A segmentation model trained with report-, morphology- and anatomy-aware supervision traces the lesion, including where it extends past the pancreas.
Review the edge
See the most likely boundary, plus a graded band where it's uncertain. It supports the radiologist's delineation. It is not a diagnosis, and the radiologist stays in charge.
Pricing
Pay per scan. Nothing else.
No seats, no licence fees, no minimums. A radiology team pays for the scans it contours, and nothing when it doesn't.
At launch
Per scan
One CT in, one tumour contour and uncertainty band out.
- Upload a CT and get the highlighted tumour back
- Pay only for scans processed: no subscription
- Built for radiologists reading abdominal CT, starting with pancreatic cancer
- Pilot pricing for early partner teams
Why the edge
Finding a tumour isn't the same as outlining it.
Pancreatic cancer is infiltrative: it wraps vessels and often extends past the gland. Expert hand-drawn outlines exist for only a few thousand public scans, while hospitals hold far more CTs paired with written reports.
Methods that learn from reports turn a reported size into a ball-shaped target, which works for finding tumours. They then cut away anything predicted outside the pancreas, which removes real tumour at exactly the edge that matters.
Ball target + hard cut
The target ignores the tumour's real shape, and tissue outside the organ mask is deleted.
Soft contour + soft boundary
Supervision that follows the reported margin, and an anatomical prior that guides the model instead of amputating tumour.
Evidence so far
Measured, cited, including the nulls.
Every number below comes from our own experiments, labelled with the study and cohort it came from. We don't have clinical accuracy figures yet, and we won't invent them.
7of 8
development cases with tumour had reference tumour outside the supplied pancreas mask: 74% of reference volume. A hard organ cut would delete it.
Task 4 · E01 · 16 native development cases · saved outputs, no retraining
14.6%
of all reference tumour voxels fall outside the anatomy search region (expert anatomy, 26 of 143 cases). With predicted anatomy: 3.92%, 17 cases.
Task 4 · S1 · 143 training cases · unchanged baseline dilation, 1 mm crops
−0.0007
mean Dice change from simply softening the search region by 5 mm. That's far short of the +0.020 we pre-registered, so we're reporting it as a null.
Task 4 · S2 · 143 exposed cases · target diagnostic, no training
How to read these. They describe the problem on development cases the team has already seen. They are not held-out performance, and they are not a claim that our method beats anything yet. The research page has the figures, caveats and what we'll count as a real win.
Roadmap
Where the product is headed.
From one model to the tool radiologists open for every tumour they read.
Now
Soft-contour engine
The model that traces a tumour's real edge, not a sphere.
Next
Scan-to-contour MVP
Upload a CT, get the tumour highlighted with its uncertainty band.
Soon
Radiology pilots
First partner teams on per-scan pricing, shaping the viewer with us.
Soon
Works in your viewer
DICOM and PACS integration, so contours open where radiologists already read.
Later
Automatic measurements
Size, volume and vessel contact from the contour, ready for the report.
Later
Follow-up tracking
Compare contours across scans to see growth or response to treatment.
Later
More tumour types
From pancreas to liver, kidney and beyond.
Vision
Clinical clearance
Validated and cleared for clinical use: the contour every radiologist trusts.
Team
Eight builders, one edge.
One team with one goal: show radiologists where a tumour really ends. Engineers and researchers building in Sydney.
Meet the teamRadiologists wanted
Tell us how uncertainty should look.
If you read abdominal CT, twenty minutes of your opinion on how a contour and its confidence band should be shown would shape what we build next.