AI-powered radiology worklist & dictation cockpit — a unified workspace where radiologists manage their reading queue, view DICOM studies, dictate reports, and let AI draft the structured report — integrated with clinical systems via HL7 and DICOM.
Cockpit de worklist e ditado radiológico com IA — um espaço único onde radiologistas gerenciam a fila de leitura, visualizam estudos DICOM, ditam laudos e deixam a IA estruturar o rascunho — integrado a sistemas clínicos via HL7 e DICOM.
⚠️ Demo project / Projeto de demonstração. Synthetic patients and generated DICOM only. No real PHI, ever. / Somente pacientes sintéticos e DICOM gerado. Nenhum PHI real.
flowchart TB
web["React/Vite cockpit<br/>worklist · dictation · admin KPIs"]
gw["api-gateway (NestJS)<br/>auth · RBAC · 403 audit · WS fan-out"]
web -- "HTTPS + WebSocket (JWT)" --> gw
subgraph services["Event-driven microservices"]
wl["worklist-svc (NestJS + PG)<br/>queue · SLA · claim/sign<br/>audit · stats"]
integ["integration-svc (NestJS)<br/>HL7 ORM in (MLLP) · ORU out<br/>Orthanc bridge · synthetic DICOM"]
dict["dictation-svc (NestJS + PG)<br/>report draft → signed<br/>signing saga with worklist"]
ai["report-ai-svc (Python/FastAPI)<br/>LLM draft · critical findings<br/>eval harness (Anthropic/OpenAI/stub)"]
end
gw -- "HTTP (X-User-*, traceparent)" --> wl & integ & dict
dict -- "draft (HTTP)" --> ai
dict -- "signing saga (HTTP)" --> wl
nats[("NATS JetStream<br/>outbox → relay → durable consumers → DLQ")]
wl <--> nats
integ <--> nats
dict <--> nats
nats -- "events → WS" --> gw
subgraph infra["Infrastructure"]
pg[("PostgreSQL<br/>database per service")]
pacs["Orthanc + OHIF<br/>PACS / viewer"]
obs["OTel → Jaeger<br/>Prometheus → Grafana"]
end
wl -.-> pg
dict -.-> pg
integ -.-> pacs
Key patterns / Padrões centrais: rich domain aggregates (DDD tático), transactional outbox,
optimistic locking, durable consumers with DLQ, append-only audit log (DB trigger), JWT + RBAC
at the gateway, LLM eval harness with CI gate. Decisions are recorded in specs/
(ADRs 0001–0009).
NestJS/TypeScript · React 19 + Vite · PostgreSQL 16 · NATS JetStream · Python 3.12/FastAPI · Docker Compose · OpenTelemetry + Jaeger · Prometheus + Grafana · Orthanc + OHIF
# full stack (13 containers)
docker compose up -d --build
# optional: real AI drafts (otherwise a deterministic stub is used)
AI_PROVIDER=anthropic ANTHROPIC_API_KEY=sk-... docker compose up -d report-ai| Surface | URL | Credentials |
|---|---|---|
| Cockpit (web) | http://localhost:5173 | ana/ana (radiologist) · admin/admin · tech/tech |
| API gateway | http://localhost:3010/api/v1 | JWT via POST /auth/login |
| Orthanc + OHIF | http://localhost:8042 | — |
| Jaeger (traces) | http://localhost:16686 | — |
| Grafana (dashboards RadFlow) | http://localhost:3300 | admin/admin |
| Prometheus | http://localhost:9091 | — |
Feed it with synthetic HL7 orders (each one becomes a study + DICOM images in Orthanc):
pnpm --filter @radflow/integration run feeder -- --rate 20 --duration 60pnpm install
pnpm run build:packages # shared → messaging → ddd → telemetry
pnpm run typecheck
pnpm test # unit + integration (integration uses Testcontainers: PG, NATS, Orthanc)
pnpm -r run test:e2e # service-level e2ePython service:
cd services/report-ai
uv sync
PYTHONPATH=src AI_PROVIDER=stub uv run pytest
PYTHONPATH=src AI_PROVIDER=stub uv run python -m report_ai.eval # LLM eval harness (CI gate)Engineering rules live in AGENTS.md. Tests are split into .spec.ts
(unit), .int-spec.ts (Testcontainers) and .e2e-spec.ts.
- Append-only audit log per service — a database trigger rejects
UPDATE/DELETE; every write use case records who/what/when/origin in the same transaction. - Denied write attempts (403 at the gateway) are audited too.
- RBAC per route:
technologistfeeds orders,radiologistclaims/dictates/signs,adminreads audit trails and KPIs. - Zero PHI: synthetic patients, generated DICOM, no audio ever leaves the browser (Web Speech API — see ADR 0006).