Seven validated steps, then the myocardium
Each step hands an inspectable result to the next: masks, centerlines, named vessels, lumen contours, FFR values and the myocardial map. A clinician can check every intermediate on screen.
Heart & coronary segmentation
Coronary lumen, aorta, LV myocardium, LV blood pool and the rest of the heart in one pass.
Dice 0.84 coronary · 0.93 LVCoronary ostia
Left and right coronary origins located on the aortic root.
0 of 42 ostia missedCenterline tracking
Network tracker trained on expert centerlines follows every branch from the ostia.
main-vessel recall 0.93Artery labelling
Tree model names LAD, CX, RCA and their side branches.
territory correct 97–99 %Lumen contours
Cross-sectional lumen at every centerline point, from nine reformatted planes.
radius error 0.14 mmStenosis grading
Diameter stenosis read from the contours along each vessel.
specificity 0.97 at ≥ 50 %CT-FFR
Tree network, monotone along the vessel, predicts FFR at every node.
vessel accuracy 0.81–0.82Myocardial ischaemia
FFR, lesions and myocardial CT mapped onto the LV in AHA 17-segment space. Trained on paired CT perfusion; growing with data.
Held-out results
Every number below comes from a fixed test set of 21 patients that was never used for training, model selection or tuning. The reference is a full expert CT-FFR analysis: centerlines, lumen contours, marked stenoses and FFR on every node, 104,537 nodes across 129 patients in total.
Segmentation
| Metric | Result |
|---|---|
| Coronary lumen Dice | 0.841 |
| Coronary HD95 (mm) | 4.4 |
| Connectivity (clDice) | 0.833 |
| Proximal 40 mm coverage L / R | 0.96 / 0.97 |
| LV myocardium Dice | 0.930 |
| LV blood pool · aorta · heart Dice | 0.956 · 0.968 · 0.960 |
Mean over 21 patients.
CT-FFR
| Input lumen | Expert | Automatic |
|---|---|---|
| Point-wise FFR error (MAE) | 0.046 | 0.060 |
| Correlation with reference (r) | 0.91 | 0.85 |
| Lesion AUC | 0.946 | 0.865 |
| Lesion sensitivity / specificity | 0.93 / 0.65 | 0.85 / 0.75 |
| Lesion accuracy | 0.85 | 0.82 |
| Vessel AUC | 0.909 | 0.884 |
| Vessel sensitivity / specificity | 0.84 / 0.81 | 0.80 / 0.81 |
39,311 centerline points, 66 lesions (46 with FFR ≤ 0.80), 175 vessel ends. "Automatic" uses our own lumen contours.
Lumen contours and stenosis
| Metric | Result |
|---|---|
| Contour Dice | 0.858 |
| Dice inside stenoses | 0.861 |
| Radius error (mm) | 0.141 |
| Diameter-stenosis error | 0.098 |
| ≥ 50 % stenosis: specificity | 0.97 |
17,226 expert contours. Results unchanged with centerline offsets up to 0.5 mm and tilts up to 10°.
Ostia, tracking and labelling
| Metric | Result |
|---|---|
| Left ostium: median error · within 5 mm | 2.9 mm · 100 % |
| Right ostium: median error · within 10 mm | 3.6 mm · 95 % |
| Tracking recall, main vessels | 0.93 |
| RCA / LAD / CX recall | 0.94 / 0.91 / 0.90 |
| Tracking precision | 0.80 |
| Territory labelling (LAD / CX / RCA) | 0.97–0.99 |
| Left / right tree assignment | 100 % of 93 |
Tracking: 10 held-out patients, tolerance 1.5 mm, against expert centerlines.
From coronary physiology to the myocardium
The final stage predicts where the myocardium is under-perfused, from the CCTA alone. It is trained against quantitative stress CT perfusion, mapped patient by patient onto the CCTA, and it already runs inside the software.
- 0.88median myocardial Dice of the CT perfusion to CCTA registration (IQR 0.86–0.90), 92 % median LV coverage
- AHA 17targets live in an LV-intrinsic coordinate system (apex to base, circumference, wall depth), so CT perfusion, SPECT and PET all train the same model
- 233patient CCTA + perfusion dataset processed, further 500 secured, with a CCTA-to-SPECT patient mapping, ongoing processing
Function from the scan you already have
- Lesion-specific CT-FFR at every point of the tree, so the haemodynamic relevance of each stenosis is visible next to its anatomy.
- Named vessels and territories: LAD, CX and RCA are labelled automatically and linked to their myocardial territory.
- Explainable chain: masks, centerlines, contours and FFR values are all on screen and can be checked and corrected.
- No stress agent, no extra radiation: the analysis uses the routine CCTA only.
- Minutes on a workstation: segmentation 4–7 min, everything after it about 2 min on a laptop GPU.
Own models, own data pipeline, protected method
- Every model is our own: segmentation, ostia, tracking, labelling, lumen, FFR and ischaemia, trained on expert-annotated data. No third-party segmentation model remains in the product.
- Strict validation: a fixed held-out test set, patient-level splits and no tuning on test patients.
- Modality-independent targets: perfusion is learned in AHA 17-segment space, so each new perfusion dataset adds to the same model.
- Intellectual property: international patent application PCT/IB2023/059985, covering multi-stage anatomical inference, physiological territory mapping and voxel-wise ischaemia estimation.
The anatomy-to-FFR chain is built. Paired perfusion data is the next step.
Every step from the raw CT to CT-FFR is trained and validated on held-out patients. The ischaemia stage is the one limited by data volume: with about 120 perfusion patients, each coronary territory has only a few dozen ischaemic examples. Scaling to several hundred paired CCTA and perfusion patients is the direct path to a production ischaemia output.
What the next round funds
- Multi-centre paired CCTA with stress perfusion (CT perfusion, SPECT, PET), building on the secured 233-patient CCTA + SPECT dataset.
- Prospective clinical validation of CT-FFR and ischaemia against invasive FFR and perfusion imaging.
- MDR and FDA clearance, and the first reimbursed deployments.
We are preparing a USD 3M pre-Series A round.
Talk to us about the round