Application Note: On-Premises Clinical Imaging AI with DICOM and MONAI

Data path

1. A DICOM router sends allowlisted studies to an isolated ingress.
2. The service validates modality, series completeness and required tags.
3. Transient pixel data is staged to encrypted local NVMe.
4. A versioned MONAI application performs preprocessing and inference.
5. Outputs are encoded as approved DICOM objects or a review record.
6. A separate authorised integration returns results after policy checks.

Platform placement

AxiRelay runs several smaller modality models or replicas. AxiAnvil fits a large 96GB model in one device. AxiCrucible separates validation from operational service and supplies node-level maintenance flexibility. Size against full study size, simultaneous arrivals and p99 latency, not one image.

Model package

Pin the model, transforms, label map, framework, container digest and supported input contract. MONAI Deploy demonstrates DICOM-aware application construction; adapt it to local routing and audit requirements. Reject unsupported series explicitly rather than resizing or coercing silently.

Validation

Evaluate by scanner, protocol, site and population; include missing/corrupted series and out-of-distribution examples. Compare shadow outputs with expert review. Measure sensitivity/specificity or task-appropriate metrics plus calibration, latency, failure rate and clinician disposition.

Safety and operations

Keep a downtime route, health monitoring and one-click rollback. Do not log patient pixels or identifiers in ordinary service metrics. Every model release needs risk review, acceptance evidence and a defined owner. Hyperion provides compute and on-prem control; it does not itself confer medical-device approval.