Radiant runs inference on each CT study, sorts the radiologist's worklist by clinical urgency, shows where the model looked and how its findings agree with the patient's clinical picture, then syncs a structured result into your PACS and EHR. The radiologist stays accountable — this is a triage decision-support aid, not a diagnostic device.
In acute care, a late or missed time-critical abdominal finding is measured in patient harm and liability — yet radiologists read against an unprioritised queue, and a bare probability tells them little about how urgent a study really is.
The Amplinth loop, tuned to radiology triage. A deterministic rules engine — not the language model — owns every clinical decision, so each result is auditable and defensible.
Six-stage pipeline reconstructs the CT and runs Pillar-0 detection — 20 calibrated finding probabilities plus attention.
A deterministic rules engine assigns an urgency tier (RED / ORANGE / YELLOW / GREEN) and re-orders the worklist.
Attention overlays and key slices show where the model looked; probabilities are calibrated to real prevalence.
Findings checked against vitals, labs & history for concordance; the radiologist confirms, disagrees or amends.
Standards-based output to PACS/EHR; value framed as triage throughput & turnaround.
Measured via EVA™Radiant is a focused clinical-AI product for one modality and body region — abdomen-pelvis CT in the acute setting. It complements PACS and reporting systems rather than replacing them.
Preprocess → Pillar-0 detection → MedSAM2 localization → MedGemma grounding → deterministic rules engine → LLM narrative. Each stage persists before the next, so the pipeline is resumable and every step is recorded.
The worklist returns studies sorted by urgency tier by default, with live Server-Sent-Events updates and per-study pipeline visibility — the most time-critical studies surface first.
Per-finding Score-CAM / Grad-CAM attention heatmaps and key-slice ranking, rendered directly in an OHIF / Cornerstone3D CT viewer — no PNG export, no viewer swap. Probabilities are isotonically calibrated to observed prevalence.
For each positive finding, the rules engine compares the AI call to the patient's vitals, labs and history and returns concordant / discordant / unassessable with auditable, non-LLM reasoning. Never-miss findings use lowered thresholds for sensitivity.
Primary finding with anatomy, measurements and evidence; related sub-threshold findings; a collapsible list of cleared differentials; and a recommended communication tier — beside the CT viewer with attention and segmentation overlays.
Mark reviewed, disagree, amend or escalate. Amendments propagate automatically into every export; escalations push a live worklist update. A full per-study review history feeds clinical-quality reporting.
DICOM-SR, DICOM-SEG and GSPS to PACS; a FHIR R4B DiagnosticReport to the EHR; DICOMweb retrieval, HMAC-signed webhooks and a PDF narrative — so downstream systems stay in sync without re-entry.
JWT auth with role-based access (radiologist / admin / integrator), scoped API keys, model-version and threshold management with never-miss safeguards, and an append-only audit log enforced by both a DB trigger and privilege separation.
20-finding abdomen-pelvis taxonomy — active hemorrhage, ruptured AAA, pneumoperitoneum, mesenteric ischemia, appendicitis, cholecystitis, bowel obstruction and more — each with tier, SNOMED code, sex-specificity, threshold and never-miss flag as code-level clinical-safety invariants. DIMSE C-STORE push ingestion and multi-tenant SaaS are on the near-term roadmap (coming soon).
Radiant runs end-to-end with a full radiologist and admin UI and a deployed Pillar-0 inference endpoint. It is not an FDA-cleared / UKCA / CE-marked diagnostic device, and every model-performance figure below is a validation / held-out-test result — not a realized clinical outcome.
The value is faster, better-ordered triage: time-critical studies surface first, with evidence and clinical context side by side.
Engineering rigour: ~25 registered FastAPI routers; a six-stage resumable pipeline with a structured error taxonomy; CI standards targeting 80% coverage on services/rules plus integration and Playwright E2E tests. Calibration: a reported probability matches observed prevalence post-isotonic — honest values, not overconfident raw scores. No hospital, site or client is named; reference material is illustrative only.
Radiant runs as SaaS on Amplinth's AWS, in any cloud, or as a single-tenant customer-VPC / on-prem appliance — so PHI can stay inside your estate. It slots into existing PACS/EHR over open standards, with the radiologist accountable at every step. Regulatory clearance (FDA 510(k) / UKCA / CE) is a separate workstream and not in place; Radiant is positioned as a clinical decision-support / triage aid.
A deterministic rules engine owns tier and concordance with a recorded rules_fired list; the language model writes prose only and can never override it. Every result is auditable to the prompt.
Attention overlays and key slices show where the model looked, in your own viewer. Calibrated probabilities mean a reported number matches the real-world rate — honesty radiologists can trust.
DICOM-SR/SEG/GSPS, FHIR R4B and DICOMweb let Radiant slot into your existing PACS/EHR estate. The radiologist remains the decision-maker — always.
Book a walkthrough on de-identified demo cases: the tier-sorted worklist, in-viewer attention overlays, deterministic clinical concordance, and one-click standards-based export to PACS and the EHR.