Orchid
An open-source radiology AI orchestration platform. Ingest over DICOM, route to AI apps by rule, collect structured results, validate against your own ground truth.
What is Orchid? Orchid is a free, MIT-licensed, open-source platform that receives radiology studies over standard DICOM, routes them to AI applications by rule, returns findings as structured results and DICOM Structured Reports, and validates every model against your own ground-truth data — a self-hostable alternative to commercial radiology AI orchestration platforms, built by 5C Network's founder.
Not a medical device — read this first
Orchid is a young project intended for research, evaluation and integration prototyping. It is not a medical device, holds no regulatory clearance anywhere, and ships with demo models that are for illustration only — not for clinical use. Do not put it in the path of patient care.
What Orchid is
Orchid sits between an imaging network and AI vendors. It listens for studies over standard DICOM — a real C-STORE endpoint any modality, PACS or router can point at — decides by rule which AI applications should see each study, calls them, and collects the findings back as structured results and DICOM Structured Reports. Around that loop it adds the tooling deployments actually need: a marketplace for registering apps by manifest, a viewer that renders AI overlays, monitoring of volume, latency and positivity rates, and a validation workbench that measures any connected model against ground-truth datasets built from your own studies, with AUC, ROC curves and confusion matrices. It is open, MIT-licensed, and self-hostable — the whole loop runs on your own infrastructure.
What it does
Everything below is in the repository today — no roadmap items are listed here.
| Capability | What it does |
|---|---|
| DICOM-native ingestion | A real C-STORE listener that any modality, PACS or DICOM router can send studies to directly, with HTTP upload as an alternative path. |
| AI app marketplace | AI applications register with a manifest declaring their task, modality, body part and endpoint. Install or uninstall, enable or disable, and test connectivity from the console. |
| Rule-based routing | Priority-ordered rules match on modality, body part, source AE title and study description; every matching app receives an inference job for the study. |
| Orchestration engine | An asynchronous worker pool processes jobs with retries, latency tracking and a per-job audit record. Studies settle, route, process and complete. |
| Structured results & DICOM SR | Findings — classifications, detection boxes, measurements — are stored as structured JSON and emitted as DICOM Basic Text SR, optionally auto-forwarded to a destination PACS node. |
| Built-in viewer | A web study viewer that renders AI overlays, with detection boxes drawn onto the images server-side per result. |
| Model validation | Build ground-truth datasets from your own studies, run any registered app across them, and get AUC, ROC curves, sensitivity/specificity, confusion matrices and per-case drill-down. |
| Monitoring & audit | Per-app volume, latency and positivity-rate tracking, plus a full audit trail of every platform event. |
Quick start
Requires Python 3.11+ and Node 18+. The demo generates a synthetic imaging cohort, sends it over real DICOM, watches the routing rules fire, and runs a validation study per app — end to end in a few minutes.
git clone https://github.com/kalsiva/orchid.git
cd orchid
make setup # venv + backend deps + web build
make dev # API (:8000), DICOM listener (:11112), demo AI apps
make demo # full end-to-end demo flow (second terminal)
Then open http://127.0.0.1:8000 for the dashboard, worklist, marketplace and validation. Prefer containers? docker compose up --build works too.
Who it's for
Researchers
A validation workbench you can point at any model behind an HTTP endpoint: ground-truth datasets from your own studies, AUC and ROC out of the box, per-case drill-down for error analysis.
Imaging-IT teams
Prototype how AI routing would work on your network — real C-STORE ingestion, routing rules against your modality and body-part mix, SR results back to a PACS — before any procurement conversation.
AI vendors
A neutral harness for testing deployment integration: implement one HTTP endpoint, register a manifest, and see how your app behaves inside a DICOM-driven orchestration loop.
The relationship to 5C, stated plainly
Orchid was built by Kalyan Sivasailam, 5C Network's founder, as an independent open-source project. It is not a 5C product, and 5C does not offer commercial support for it — support is community-based, through GitHub issues.
The ideas in Orchid — DICOM-native ingestion, rule-based orchestration, structured results, relentless model validation — are the same ideas that run at commercial grade inside 5C's platform, where they power an operation reading thousands of studies a day under radiologist sign-off. If you want the open, self-hostable version to study and experiment with, Orchid is on GitHub. If you want the production system with clinical AI, radiologists and accountability built in, that is Bionic Radiology. For the wider open-source landscape — viewers, PACS servers, training frameworks — see our guide to open-source radiology AI tools.
Frequently asked questions
Is Orchid free to use?
Yes. Orchid is open source under the MIT licence, one of the most permissive licences available — you can use, modify, self-host and redistribute it, including commercially, at no cost. The bundled demo models and synthetic data are for demonstration only and are explicitly not for clinical use.
Can Orchid read scans by itself?
No. Orchid is orchestration, not diagnosis. It receives studies over DICOM, routes them to whichever AI applications you connect, collects their findings, and helps you validate those applications against ground truth. It ships with simple demo models for illustration only. Orchid is not a medical device and makes no diagnostic claims — any clinical interpretation still requires appropriately cleared tools and a qualified radiologist.
How is Orchid different from commercial platforms like CARPL?
Commercial radiology AI platforms such as CARPL.ai offer a similar shape of product — validate, deploy and monitor third-party AI through one integration — as a supported enterprise service with a curated vendor marketplace and regulatory posture to match. Orchid is a young, MIT-licensed, self-hostable open-source implementation of the same ideas: DICOM ingestion, rule-based routing, structured results and ground-truth validation, with the code fully inspectable. It offers no commercial support, no vendor catalogue and no clearances — the fair summary is that Orchid is for research, evaluation and integration prototyping, while commercial platforms are for supported production deployment.
Does 5C Network commercially support Orchid?
No. Orchid is an open-source project by 5C Network's founder, and support is community-based through GitHub issues and pull requests — there is no SLA, paid support tier or commercial offering around it. 5C's commercial platform is a separate, production system; the two share ideas, not a support contract.
Can Orchid run fully on-premise, without internet access?
Yes. Orchid is self-hostable by design: a single Python process (API, DICOM listener and orchestrator) with a statically served web console, using SQLite by default or Postgres via configuration, run directly or with Docker Compose. Nothing about the platform requires an external service, so it can operate inside an isolated hospital or lab network — the only network calls it makes are to the AI application endpoints you register.
Clone it, break it, send a PR
Orchid is MIT-licensed and contributions are welcome. And if what you actually need is radiology reporting at production scale, we should talk.
