Radion gives radiology agents calibrated instruments—and lets radiologists watch every slice they touch.
An open-source DICOM reading room, 42-tool Model Context Protocol harness, and 185-task CT, radiograph, and MR scoreboard. Code computes the millimetres and Hounsfield units; the agent cannot report a measurement it did not take.
Pulmonary nodule in the left lower lobe superior segment: 12.4 × 9.8 mm on axial slice z=64 (-16.5 mm), mean attenuation +34.2 HU, volume 1,001 mm³ [preliminary — awaiting named physician review].
Exports: DICOM SEG · DICOM TID 1500 SR · FHIR R5 DiagnosticReport
git clone https://github.com/cYoren/radion && cd radion
cp .env.example .env
sed -i "s/^RADION_API_KEY=.*/RADION_API_KEY=$(openssl rand -hex 32)/" .env
./scripts/demo.sh # Orthanc + OHIF + Reading Room on 127.0.0.1:8100Why radiology AI fails on raw screenshots—and how a deterministic harness fixes the contract.
Sending uncalibrated JPEGs to a multimodal model loses voxel spacing, rescale intercepts, and 3D orientation. Radion separates visual reasoning from physical measurement so every number in a draft report traces back to a verifiable DICOM operation.
| Clinical & Engineering Requirement | Uninstrumented Vision-Language Prompting | Radion Harness + Reading Room |
|---|---|---|
| Linear, RECIST & Cobb geometry | Estimated by eye from rescaled 2D pixels; ignores anisotropic slice thickness | Computed in 3D patient LPS millimetres from affine cosines; 4-point RECIST checks 90° ± 10° orthogonality |
| Attenuation & density claims | Windowed grey levels confused with Hounsfield Units | Calibrated RescaleSlope · SV + RescaleIntercept before any ROI or threshold tool executes |
| Hallucinated report numbers | Free-text narrative can invent sizes, slices, or series numbers | propose_finding and draft_report reject any number not backed by a ledgered measurement_id |
| Radiologist supervision | Opaque final paragraph; impossible to see which slices the model actually inspected | Live SSE step stream, animated FOV crop box, visited-slice rail, and 1-key accept / correct / reject |
| Archive interoperability | Markdown blob trapped in a chat window | Writes standards-compliant DICOM SEG, DICOM TID 1500 SR, and FHIR R5 DiagnosticReport back to Orthanc / OHIF |
| Negative-screen phrasing | Over-confident “exam is normal” statements on narrow triage heads | Explicitly lists checked vs unchecked sections; never emits “normal” as a clinical rule-out |
42 MCP tools across 8 radiological toolsets.
Both the radiologist's hand instruments on the Reading Room canvas and external AI agents over stdio or http://127.0.0.1:8812/mcp call the exact same Python functions. Every call is recorded to an append-only ledger with arguments, latency, view artifacts, and provenance digests.
- Enhanced Multi-frame & NIfTI-1 ready
- Expands
PerFrameFunctionalGroupsSequencevolumes, targets individual series viaseries_uid, and imports research.nii / .nii.gzwith automaticfrom_dicom=FalseSEG/SR guards. - ADR 0014 PHI data-class lock
- Uploaded clinical exams default to
restricted. Remote model calls are refused unless the study ispublicor carries a nameddeidentified_attestedrecord. - Corroborated multi-tool evidence
- Independent detectors fuse into content-addressed evidence bundles graded A–D with recorded conflicts, skip reasons, and patient-bootstrap confidence intervals.
core
Study navigation, affine coordinates & calibrated viewslist_studies, get_study, list_series, get_acquisition, find_landmarks, find_extreme, map_point, render_view, mark
measure
Voxel-to-millimetre & Hounsfield physics in patient LPS spacemeasure_distance, measure_bidimensional (RECIST 1.1 4-point), measure_cobb, measure_angle, measure_area, measure_roi, measure_mask, measure_ratio, measure_profile
segment
3D connected-component & HU-bounded volumetric maskssegment_region, segment_threshold
compare & calculators
Longitudinal volume doubling & published clinical formulascompare_prior, calculate (Fleischner, Lung-RADS, volume doubling time, adrenal washout, cardiothoracic ratio, NASCET, Evans index)
quality & evidence
Slice geometry checks, corroborated detector graph & linked reportingcheck_quality, get_evidence, propose_finding (rejects unmeasured numbers), draft_report (EN / PT structured templates)
bench
Leak-tested evaluation over stdio & streamable HTTPlist_tasks, start_task, submit_task, score_run
An objective 185-task scoreboard across CT, chest radiograph, and MR.
Instruments without a benchmark are a demo; a benchmark without instruments measures guessing. Radion ships four public task packs with an oracle ceiling (feeding reference masks through the real measurement tools to prove zero-bug voxel-to-mm plumbing) and a scripted non-ML floor, plus leakage tests guaranteeing ground truth never crosses the MCP wire.
| Modality & Pack | Measured Task Families | Oracle Plumbing Check | Ground-Truth Provenance |
|---|---|---|---|
| Chest CT (Synthetic Phantom) | Planted solid & subsolid pulmonary nodules, lung volumes, LAA-950 emphysema | Analytical mask sanity check (≤ 1 voxel margin) | Deterministic generator (`radion phantom`, Apache-2.0, 2026) |
| Computed Tomography (Public Masks) | Organ volumetry, lesion longest-axis & perpendicular RECIST diameters, HU attenuation | Reference-mask voxel→mm plumbing check | Public reference masks documented in `docs/BENCH-SPEC.md` (2026) |
| Chest Radiograph (DX / CR) | Cardiothoracic ratio, projection distances & calibrated RAD-DINO triage operating points | Projection geometry + patient-bootstrap CIs (`docs/EVALUATION-CXR.md`) | MIT `microsoft/rad-dino` frozen encoder + patient-disjoint Platt calibration (2026) |
| Magnetic Resonance (MR) | Cross-sectional area, multi-series `series_uid` targeting & volumetric mask metrics | Reference-mask affine conversion check | Enhanced Multi-frame DICOM & NIfTI-1 (`from_dicom=False`, 2026) |
Benchmark provenance: reproducible reports published at docs/EVALUATION-AGENT.md, docs/EVALUATION-CXR.md, and docs/BENCH-SPEC.md (generated by radion bench report · scope: non-clinical software benchmark · source year: 2026 · last checked: 2026-11-17 · owner: Medhub Engineering, radion v0.1.1 · review cadence: 90 days). Open roadmap items include live frontier-model runs (B3), promptable anatomy segmentation adapters (B4 / T2), and labelled claim-level interpretation packs (B5).
One Docker workstation. Zero mandatory cloud dependencies.
Radion deliberately avoids microservice sprawl and proprietary viewer lock-in. It composes proven open-source medical imaging infrastructure behind a single loopback gateway so imaging departments, teleradiology groups, and AI labs can evaluate it on a single Linux or macOS machine in minutes.
- Orthanc 1.13 + OHIF 3.13: standard DICOMweb (QIDO-RS, WADO-RS, STOW-RS) and C-STORE port 4242, with automatic background polling for newly arrived stable studies.
- Postgres 17 + Content-Addressed Store: durable
FOR UPDATE SKIP LOCKEDjob queue, immutable SHA-256 mask artifacts, unified benchmark/study ledger, and physician review audit trail. - Permissive Apache-2.0 Core: code licences and neural weight licences are audited separately in
LICENSE_MATRIX.md; copyleft archive components sit strictly across protocol boundaries.
Evaluate Radion with your imaging team or AI lab
Run the full Apache-2.0 stack yourself today, or scope a focused technical pilot with Medhub for private PACS integration, custom structured report templates (EN / PT / DE / FR / IT), or site-specific benchmark packs.
- On-premise / VPC deployment alongside Orthanc or hospital PACS
- Benchmarking internal or open-weight VLMs against your own reviewed cases
- Bilingual structured reporting workflows with SOP-level traceability
Frequently asked questions
Why do vision-language models need a deterministic DICOM harness instead of reading PNG screenshots?
A PNG screenshot strips pixel spacing, slice thickness, rescale slope/intercept, patient orientation cosines, and FrameOfReferenceUID. When an LLM guesses '12 mm' or '35 HU' from pixels alone, the number is uncalibrated text. In Radion, Python computes every distance, angle, area, volume, and Hounsfield statistic directly from validated DICOM or NIfTI voxel geometry, and propose_finding rejects any numeric claim that lacks a recorded measurement_id.
How does the Reading Room integrate with OHIF Viewer and an existing PACS?
Radion ships with Orthanc 1.13 (DICOMweb + C-STORE on port 4242) and OHIF 3.13 out of the box. The Study screen defaults to the Live Reading Viewport—showing the agent's animated field-of-view crop, visited-slice rail, and click-to-measure calipers—and provides a 1-click toggle into OHIF hydrated with exported DICOM SEG masks and DICOM TID 1500 Structured Reports.
Does patient imaging data ever leave the hospital or workstation?
No by default. Every Docker Compose port binds to 127.0.0.1, and every study carries an ADR 0014 data class (public, deidentified_attested, or restricted). Uploaded clinical DICOMs default to restricted: remote LLM APIs and external MCP egress are blocked in code unless an operator explicitly configures local endpoints or records a named de-identification attestation.
Which AI agents and MCP clients work with Radion?
Any client speaking the Model Context Protocol (stdio or streamable HTTP on :8812/mcp) can drive Radion's 42 tools—including Claude Code, Claude Desktop, custom Python/TypeScript MCP agents, or Radion's built-in OpenAI-compatible multimodal runner (which supports local vLLM, Ollama, llama.cpp, or gated cloud endpoints). External MCP tool calls targeting a study automatically stream live into the Reading Room over Server-Sent Events.
How do radiologist corrections improve future evaluations without leaking PHI?
When a radiologist clicks Accept, Correct, or Reject on a proposed finding in the Reading Room (or re-measures a lesion with the hand caliper), Radion records the named physician review, updates the linked EN/PT structured report, and compiles accepted or corrected measurements into a local private benchmark pack (site-reviewed) containing study UIDs and millimetre coordinates with zero patient identifiers.
Is Radion a cleared diagnostic medical device?
No. Radion is open-source Apache-2.0 software built for research, objective agent benchmarking, and physician-supervised clinical decision support. It is not cleared, certified, or registered by the FDA, EMA, MHRA, or ANVISA, never emits the word 'normal' as a rule-out, and embeds its research disclaimer as a structured field on every report.
Research, benchmarking, and physician-supervised decision support boundary
Radion is open-source engineering software (Apache-2.0) and is not a certified or cleared medical device under FDA, EU MDR, UKCA, or ANVISA regulations. Shipped detectors and benchmark baselines are not clinically validated for autonomous diagnosis, screening rule-out, or unattended patient triage. Every generated report carries this notice as a required structured field and requires review and sign-off by a qualified physician.