0.92DXWrist · pediatric
Distal radius fracture0.92
Buckle fracture · saliency-confirmed on bone
// YOLOv8 · GRAZPEDWRI-DX
// pacs-native radiology AI
See the fracture in 8 seconds.
Not the next slot.
Madad reads every study the moment it lands in PACS — surfaces what's urgent, drafts the report, writes findings back to PACS and EHR. Your radiologists never leave their viewer.
distal radius fracture · 0.92distal radius fracture
YOLOv8 · saliency-confirmed
// the difference
Same X-ray. Two timelines.
One pediatric wrist. One distal radius fracture. The image hasn't changed in fifty years. What changes is everything around it.
without madad≈ 52 min · 7 hops
01
Tech
captures
captures
02
Sits in
queue
queue
03
Rad opens
blank report
blank report
04
Rad reads<br/>+ marks
05
Manual
SR write
SR write
06
Manual
EHR push
EHR push
avg time-on-task: 5–7 min per studyqueue-position lotteryurgent ones buried
with madad≈ 90 sec · 3 hops
01
Tech
captures
captures
02
Madad triages
in 8 sec
in 8 sec
03
0.92Rad confirms<br/>AI finding
04
Auto SR
+ FHIR push
+ FHIR push
8s to AI triagedraft report readyurgent at top of queue
// figures based on internal time-on-task observations against the M1 wrist fixture (Wikimedia public-domain buckle fracture).
// what it sees
Real images. Real outputs.
Every confidence number below comes from the deployed model running on a public-domain X-ray bundled with the repo — including the case where the model honestly misses, and flags for review.
0.92DX
CR
CR
CR// image sources: Wikimedia Commons public domain · model outputs reproducible end-to-end on a fresh `docker compose up`.
// the loop
Four steps. One trace ID.
Every transition is captured. Replay any study end-to-end in the audit viewer.
01
DICOM lands
Orthanc / your PACS webhooks Madad on every stable study. PHI redacted at ingest.
02
AI reads
Right model for the body part — YOLOv8 wrist, DenseNet CXR. Saliency + calibrated confidence.
03
Rad confirms
Study opens in OHIF with bbox, finding cards, and a draft report. Accept, reject, edit.
04
Truth flows back
Findings written as DICOM SR to PACS and FHIR DiagnosticReport to EHR. Trace ID end to end.
// trust by design
PHI never reaches the model.
Patient name, MRN, accession, dates, institution — stripped at ingest using DICOM PS3.15 Annex E, before the model sees a single pixel. Every step has a trace ID, every prediction has a registry entry, every dataset has a license row.
PACS
Orthanc · DICOMweb
Annex E redact
PHI strip · ingest
AI inference
YOLOv8 · DenseNet
PACS + EHR
DICOM SR · FHIR DR
discarded
never persisted
audit log
every step · trace_id
Every step, traceable
One trace ID follows a study from PACS arrival to FHIR push. The audit viewer is filterable by service, action, role, study UID, and time range.
Know which weights ran
Every body part has a registry entry — weights SHA-256, activation timestamp, who approved it. Hot-swap with a queue-drain. Full activation history.
Every input, accounted for
One CSV register lists every dataset, model, and dependency that touches the system — with separate columns for case-study and commercial use.
// where it's going
Wrist X-rays today. Imaging-AI substrate tomorrow.
Madad's V1 reads pediatric wrist fractures and chest X-rays. That's not the destination — that's the shape of the proof. The same stack extends to every imaging modality your hospital touches.
- ◆From a tool to a substrateThe wedge is pediatric wrist. The architecture isn't. The PACS plumbing, audit layer, model registry, and bilingual surface all generalize to any imaging modality and any pathology — adding a body part is a model entry, not a rebuild.
- ◆From a vendor to a platformMadad ships every component as something the buyer can audit, swap, or extend — open standards out, modular workers in. The platform is the product; the V1 case study is just one valid shape it can take.
- ◆From AI-assist to AI-assuredThe next phase isn't more findings — it's more confidence per finding. Federated eval, drift monitoring, calibration tracking, and per-site validation are the path from radiologist-assist to clinically-assured assist.
// roadmap
shipped
Pediatric wrist fracture + chest X-ray multi-finding
YOLOv8 trained on GRAZPEDWRI-DX for wrist; torchxrayvision DenseNet for CXR. Both with calibrated confidence and saliency overlays.
shipped
PACS + EHR round-trip, full audit, model registry
DICOM SR writeback, FHIR DiagnosticReport push, end-to-end trace IDs, PHI redaction at ingest, hot-swap model registry — all live.
in flight
Site-paced eval & drift monitoring
Re-run measured AUROC against a held-out eval set on demand, track drift over time, alert when calibration shifts on a deployed body part.
next 3 mo
Modality expansion: CT head trauma + mammography screening
Same architecture, two new model entries — drop the weights into the registry and the worklist auto-prioritizes them with the same formula.
next 6 mo
Longitudinal patient view
Surface prior reads from the EHR alongside the current study so the AI's draft report can reference baseline — not just describe a frame in isolation.
future
Federated evaluation across sites
Measure model drift and concordance across multiple hospitals without moving PHI. Each site runs the eval locally; only metrics aggregate centrally.
// pilots open
Run it on your DICOMs.
We bring the models, workers, and audit layer. You bring a PACS endpoint and a few test studies. Standalone Docker — no rip and replace.