Gathers its own context
Designed to pull priors, reports and the treatment line itself. No copy-paste, no prompts.
AI research lab
Grounded AI for oncology.
We are building toward a physician copilot for the whole cancer journey, from diagnosis to theranostics to treatment. It reads the patient's evidence like a specialist, shows where every statement comes from, and knows when to defer.
- A Liver Multiple focal lesions
- B Right lower abdomen Lesion or bowel? Flagged
- C Bladder Physiologic uptake
Why now
Answers got cheap. Reading a case did not.
Before a tumour board, someone gathers each patient's letters, results and scans by hand. In the US, PET scan volume is growing and many sites report waits of eight days or more.
+12.2%
IMV PET Market Summary, June 2025
40%
IMV PET Market Summary, June 2025
Approach
Designed the way a specialist reads.
Numbers come from tools
Every measurement is designed to be computed by code, never generated.
Evidence before conclusion
Each statement is designed to link to the record, image region or paper it rests on.
Knows when it isn't sure
Below its threshold, it is designed to flag for review instead of guessing.
How the lab works
Research and product, one loop.
We build on frontier models and own the layer above them. Physician edits are designed to close the loop.
- Grounded evidence
- Grounded perception
- Longitudinal reasoning
- Verification & calibration
- Clinical agents
- Evaluation & benchmarks
Research
New methods, datasets and open benchmarks.
Roadmap
Deeper along the patient's journey, one stage at a time.
-
Now
Grounded case summaries
A research prototype that turns patient records into source-linked tumour board summaries for clinician review.
-
Now
Imaging reasoning over time
Research on separating real change from measurement noise, so each new scan is read against the patient's history.
-
Next
Nuclear medicine reader
Designed to draft grounded reports for PET/CT and theranostics, for physician review.
-
Later · gated
Oncology copilot
Support for the treating team across diagnosis, theranostics and treatment.
-
Later · gated
Outcome-linked patient model
Aims to link imaging to outcomes: who responds, who relapses.
Later stages are gated: they start only after earlier work has proven itself with physicians.
Team
Researchers, engineers and physicians.
From leading universities, hospitals and technology companies. ViVi originated at the University of Toronto.
Earlier research
Selected publications
Earlier work by co-founder Khai Le-Duc and collaborators, across medical imaging, speech and sensing.
-
Preprint · 2025
S-Chain: Structured Visual Chain-of-Thought for Medicine
A dataset of 12,000 medical images where each reasoning step points to a region of the image.
-
Preprint · 2024
LiteGPT: Large Vision-Language Model for Joint Chest X-ray Localization and Classification
A model that finds and classifies findings in chest X-rays, evaluated on VinDr-CXR.
-
Communications Medicine · 2026
A wearable device dataset for mental health assessment using laser Doppler flowmetry and fluorescence spectroscopy sensors
A dataset from wearable sensors, built to support mental health assessment.
-
EMNLP 2025
MultiMed-ST: Large-scale Many-to-many Multilingual Medical Speech Translation
A large dataset for translating medical speech between five languages, in every direction.
-
ACL 2025 · Industry Track, Oral
Sentiment Reasoning for Healthcare
Models that explain each sentiment prediction with a written rationale.
-
Interspeech 2024 · Oral
Real-time Speech Summarization for Medical Conversations
Live summaries of doctor-patient conversations, with a new dataset, VietMed-Sum.
Our team comes from
Work with us
We collaborate across the board.
- Hospitals
- Clinicians
- Universities
- Engineers
- Industry
- Funders