Engineering / Lantern
Lantern, open discovery, governed pixels
Federated, privacy-tiered discovery for medical imaging. A researcher learns whether a cohort exists across hospitals, and the scans do not move at all.
- Challenge Brought by Boston Children’s Hospital
- Built July 25, 2026, at a healthcare hackathon in Boston
- Tests 115 passing, as the README reports
- Data Synthetic only. No real PHI
Built for a MASS AI + Open Accelerator, Boston Children’s Hospital challenge · July 2026
The question
What does the scan already know?
Pediatric and rare disease research needs cohorts that no single hospital holds. Today a researcher approaches hospitals one at a time and waits months to learn whether enough data even exists. The challenge asked how independent hospitals could make their imaging data discoverable and securely accessible without centralizing patient data.
We profiled the synthetic challenge corpus before designing anything, and one number shaped the build. In 78.4% of its reports, a quantitative clinical measurement sits locked in prose. In a real record that is a number a radiologist already measured. It is invisible to search and too risky to share, because free-text reports are the densest surface of patient information in the record.
How it works
A compiler inside the hospital, a broker outside it
- Compile inside the hospital Each hospital compiles its studies into de-identified Study Passports inside its own trust boundary. Direct identifiers are removed or pseudonymized, and ages are generalized.
- Pull the numbers out of prose Measurements become a structured, queryable axis. What crosses is the measured value and a bounded evidence snippet. The full report never leaves.
- Federate the search A broker fans a query out across three separate node services and fuses the ranked results. No model takes part in search.
- Apply disclosure policy Small cohorts are suppressed with k-anonymity, and a guard held per hospital defends against differencing.
- Petition for the pixels A request for images routes to the owning hospital with the IRB number and purpose captured, and writes an append-only audit entry.
Full DICOM is architecturally absent. Lantern does not serve pixels and never holds them, and an approval returns a simulated retrieval grant.
Who it serves
One passport, three questions
Redaction is applied on the server by role. In this prototype the role is a demo identity chosen in the console, not a verified login.
-
Computational researchers
Does a cohort exist that fits my method? Measurements are queryable by threshold, with the source sentence behind every number.
-
Physicians
Has anyone seen a case like this one? Comparable cases across institutions, and a direct route to the owning hospital.
-
Patients and families
Is my child’s condition being studied? Plain language, and counts that stop being precise before they could identify a child.
What it shows
Measured against keyword search, including a loss
The comparison runs on the synthetic challenge corpus, and its ground truth comes from Lantern’s own extractor.
-
78.4%
Measurements locked in prose
The share of reports across the synthetic network that hold a quantitative clinical measurement inside free text.
-
> 10 mm
A question keywords cannot ask
A keyword search can find a named finding. It cannot run “fetal atrial width greater than 10 mm,” because literal search has no numeric comparator. The compiled measurements can.
-
100%
Where keywords win
For pediatric brain tumor, at the one synthetic node with such studies, every literal keyword hit was also in Lantern’s expanded result set (100%), and keywords found 6.1% of that set. The write-up calls the literal search precise and simpler, and keeps the result in.
-
115
Tests reported passing
The README reports 115 passing tests on the build machine. One of them attacks the narrator with a hostile input and asserts that nothing leaks. The challenge corpus is not in the public repository, so a fresh clone runs fewer.
The limits
What Lantern does not claim
- No real patient data was used. The corpus is synthetic challenge data, on purpose.
- The de-identification is field minimization inspired by DICOM PS3.15. It is not a certified implementation, and the project does not claim HIPAA compliance.
- Extracted measurements carry stated confidence and provenance. They are not clinically validated.
- The differencing defense is tested against one-constraint subtraction. Multi-axis and cross-session attacks are named limitations.
- The three-node federation is real fan-out across three separate services, running on one laptop.
- The comparison’s ground truth is extractor-derived, not clinician-adjudicated.
- The challenge corpus is report text and metadata with no images. So there are no image embeddings, no acquisition-parameter index, and no similarity search.
Run it
Check it yourself
One command brings up the whole stack, three node services and the broker. A second checks every demo claim against the live API.
git clone https://github.com/Angiebio/healthcare-build-day-2026
cd healthcare-build-day-2026
pip install -r requirements.txt
python -m app.run_all
python tools/verify_demo.py Run the last command in another terminal while the stack is up. The console and the API are served at http://localhost:8000.
Who made this
Built on July 25, 2026 at The Open Accelerator’s Real-World Healthcare Hackathon in Boston, with Pooja Upadhyay and a cooperative of AI agents, for a challenge brought by Boston Children’s Hospital.
The institution carrying the obligation should be the institution making the call.
More builds
Other work with a page of its own
- Build Registry Metric Lab Can one extra site change the enrollment total?
- Build Continuity Ledger Can a remembered instruction outlive its permission?
- Build Adapter Promotion Gate What must hold before a model change can serve?
- Build Stagnation Observatory When does useful work look stalled?
- Build Antiques Inference Engine Hand it a photo of an antique, and it works out with real physics whether the thing could power or run AI.
- Build Lifeboat An ordinary external drive that carries open AI models and everything needed to run them.
- Build Off-Target A design change review of an AI-authored weight edit.
- Build Aurora Measures from brain science, run live on a persistent agent’s decision stream.
Get in touch
Speaking, teaching, board work, or coaching
I am glad to hear from conference organizers, journalists, students, and leaders who want a thinking partner.