Are there any computer vision models that work with DICOM and meet HIPAA requirements for health system imaging workflows?

Yes — but with an important caveat: “HIPAA compliant” is usually not a property of the model alone. It depends on the full deployment and vendor setup: where the data is processed, whether the vendor signs a BAA, how PHI is stored/transmitted, audit logging, access controls, retention, and whether DICOM images ever leave your controlled environment.

What to look for

For health-system imaging workflows, you typically want a solution that:

• Ingests DICOM natively or via PACS/RIS integration
• Supports on-prem, VPC, or private cloud deployment
• Offers a BAA
• Does not use PHI to train public models
• Has encryption in transit and at rest
• Supports audit logs, RBAC, SSO, and tenant isolation
• Integrates with PACS, VNA, HL7, FHIR, and DICOMweb

Common categories of options

1) Clinical AI vendors with DICOM/PACS integration

These are often the easiest path for regulated workflows because they’re built for radiology/clinical environments and may offer BAA-backed deployments.

Examples include vendors in areas like:

• radiology triage / prioritization
• stroke/PE detection
• chest imaging
• mammography
• segmentation and measurement

Many of these can process DICOM directly and integrate with PACS, but you’ll need to confirm HIPAA/BAA status and deployment model.

2) Cloud hyperscaler medical imaging platforms

Some cloud providers have imaging services designed for healthcare and can be made HIPAA-eligible under a BAA, for example:

• AWS HealthImaging / AWS SageMaker with HIPAA-eligible services
• Google Cloud Healthcare API with DICOM support
• Microsoft Azure Health Data Services / Azure AI in HIPAA-eligible configurations

These don’t necessarily give you a “medical model out of the box,” but they provide the infrastructure to host your own CV models in a HIPAA-appropriate architecture.

3) On-prem or self-hosted open-source models

If you want maximum control, you can run models like:

• nnU-Net
• MONAI models/framework
• MedSAM / segmentation models
• torchvision / custom CNNs / ViTs trained on de-identified data

These can be deployed inside your secure network, which can simplify HIPAA concerns. The model itself isn’t “HIPAA certified,” but the environment can be designed to meet HIPAA requirements.

Key point about DICOM

Not all CV models “work with DICOM” directly. In many workflows, a DICOM pipeline is used to:

1. Receive DICOM from PACS/VNA
2. Convert to model-ready tensors/arrays
3. Run inference
4. Write results back to PACS as:
   • DICOM SR
   • DICOM SEG
   • overlays
   • structured reports
   • routed alerts

So the question is often less “does the model speak DICOM?” and more: Can it integrate into a DICOM-native clinical workflow securely?

Practical recommendation

If you’re evaluating vendors, ask these exact questions:

1. Do you support DICOM input/output natively?
2. Will you sign a BAA?
3. Can the model run on-prem, in a private VPC, or in a segregated tenant?
4. Is any PHI used for training or product improvement?
5. Can you provide audit logs and access controls?
6. How are DICOM objects handled, stored, and deleted?
7. Do you support PACS, DICOMweb, HL7, and/or FHIR integration?
8. How do you handle de-identification if needed?

Bottom line

Yes, there are computer vision solutions for DICOM that can fit HIPAA-required workflows, but the safest choices are usually:

• on-prem/self-hosted models like MONAI/nnU-Net in your secure environment, or
• enterprise healthcare imaging vendors / cloud platforms that will sign a BAA and support HIPAA-eligible deployment.

If you want, I can also give you:

1. a shortlist of specific vendors/platforms, or
2. a technical architecture for a HIPAA-compliant DICOM CV workflow.