Reading
the invisible
signal
inside cancer.
Every scan already contains the answer. Twelve-thousand quantitative features — texture, morphology, heterogeneity, perfusion — that a radiologist, trained on human vision, can only approximate. We built CanScan to read all of them, at once, and fuse them with the molecular state of the tumor.
Read the full story →A model that sees like a tumor grows.
Cancer is hierarchical — voxel, lesion, organ, patient — and so is CanScan. Six attention stages learn structure at every scale, then fuse imaging with proteomic signal through a shared latent space. No pixel is treated in isolation; no biomarker arrives without context.
Trained on 2.1 million studies across 34 institutions, validated prospectively in four solid-tumor cohorts. Deployable as a DICOM-in / JSON-out service in under 90 seconds per study.
Lung
Screening, staging, response to ICI. Trained on 580k thoracic studies across 14 sites.
Pancreatic
The hardest cancer to catch early — we detect sub-centimeter lesions in dual-phase CT.
Renal
Non-invasive subtyping and risk stratification from contrast-enhanced CT/MRI.
Endometrial
Pre-operative risk stratification from pelvic MRI; myometrial invasion scoring.
What a scan knows about a tumor's biology.
Radiomic features encode structure; proteomic assays encode function. CanScan learns a joint embedding across both, so a CT can predict protein expression, and a blood draw can refine a scan. The result is a single, continuous readout of the tumor — not a pile of disconnected reports.
This powers three clinical jobs: earlier detection, cleaner stratification, and response prediction — the three places oncology outcomes are won or lost.
Latent space · 1,024-dim
Built with leading institutions.
Active collaborationsA platform bet on quantitative oncology.
Series A closed Q1 2026. $42M led by Polaris Bio, joined by Khosla Ventures, GV, and strategic from three top-20 pharma.
- 4 commercial indications, 2 in late-stage trial
- 11 hospital systems · 3 pharma co-development deals
- FDA 510(k) submitted Q4 2025 · CE-MDR Q2 2026
- 18 publications · 34-institution training corpus
Your scanners already saw it. Read it.
DICOM-in, JSON-out, on-prem or cloud. Integrates with PACS/RIS in under a week; radiologist workflow-native reports with confidence intervals and audit trail.
- <90s per study · on-prem GPU or HIPAA cloud
- Epic · Cerner · Sectra · Philips ISP integrations
- Structured reports · DICOM-SR · HL7 FHIR
- HIPAA · SOC 2 Type II · ISO 13485
A biomarker engine, not a black box.
Academic license for CanScan features, pre-trained encoders, and the 12,384-feature radiomic panel. Pull imaging embeddings into your own analyses.
- Academic API · 500k free studies / quarter
- Pre-trained encoders · ONNX + PyTorch
- Open feature panel · reproducible pipelines
- Collaboration desk · 34 active studies
We hire seriously curious people.
Clinicians who code. Engineers who read oncology journals for fun. Researchers who'd rather ship than paper-polish. If you want your work to matter in a scanner room in three quarters, not three decades, we should talk.