Operational dashboard
Ingest volume, queue depth, failed jobs, inference latency, activation eligibility, and hospital-specific error rates.
Scaling From Pilot To 10+ Hospitals
The system should handle site-specific integrations while keeping the internal workflow model consistent across hospitals.
Hospital Onboarding Plan
Onboarding should be a repeatable rollout workflow, not a one-off integration project. Hospital-specific setup can vary, but the internal study states and review workflow should stay consistent.
| Phase | What happens | Completion criteria |
|---|---|---|
| Setup | Register facility, supported modality scope, expected volume, integration mode, and any report status/handoff destination. | Facility config exists and required contacts/owners are known. |
| Connectivity | Configure DICOM receipt/fetch path and validate basic metadata mapping. | Test studies arrive with expected study, series, and accession references. |
| Shadow validation | Run ingestion and AI in monitor-only mode using test or shadow studies. | Display assets, AI outputs, latency, and error rates look reasonable. |
| Pilot review | Invite pilot radiologists and validate eligible-study activation, finding feedback, and optional sidecar draft workflow. | Radiologists can complete sample cases without support intervention. |
| Report status handoff | Validate how the sidecar records or receives legacy report completion status. | Completed reports in legacy systems can be correlated back to sidecar review state. |
| Expand volume | Review early dashboards before increasing study volume or adding modalities. | Operational metrics are stable enough for broader rollout. |
Operational Controls
| Control | Design |
|---|---|
| Idempotency | All ingest and export jobs use stable keys so retries do not duplicate studies or reports. |
| Dead-letter queue | Jobs that fail repeatedly are isolated with error context and visible operator actions. |
| Per-facility configuration | Feature flags, model enablement, supported modalities, templates, and handoff destinations are tenant scoped. |
| Model version tracking | Every AI finding links to the run and model version that produced it. |
| Latency SLOs | Track time from study receipt to AI-ready and from sidecar review completion to legacy report status confirmation where available. |
Dashboards
Workflow dashboard
Studies waiting for review, radiologist assignment load, review duration, AI acceptance/edit/rejection rates, and report status correlation.
Deployment Model
The initial design can support cloud-hosted, hospital-hosted, or hybrid deployment; the integration boundary should stay explicit either way.