Deploy and use MedImageInsight Premium in Foundry

Deploy MedImageInsight Premium in Microsoft Foundry and send a test request to generate medical image and text embeddings for evaluation workflows.

The healthcare AI models marked (preview) in this article are currently in limited preview. These models are intended and provided as-is for research and model development exploration. The healthcare AI models are not designed or intended to be deployed in clinical settings as-is. They are not intended for use in the diagnosis or treatment of any health or medical condition, and the individual models’ performances for such purposes have not been established.

You bear sole responsibility and liability for any use of the healthcare AI models, including verification of outputs and incorporation into any product or service intended for a medical purpose or to inform clinical decision-making, compliance with applicable healthcare laws and regulations, and obtaining any necessary clearances or approvals.

In this article, you deploy MedImageInsight Premium (preview) and send a test request to generate medical image and text embeddings. MedImageInsight Premium generates rich, semantically meaningful embeddings of medical images across nine imaging modalities, including X-ray, CT, MRI, ultrasound, dermatology, ophthalmology, pathology, mammography.

MedImageInsight Premium provides assistive model output for medical imaging workflows. Treat embeddings, similarity scores, classifications, and retrieval results as workflow inputs that still require appropriate testing, validation, monitoring, and human governance before use in clinical contexts. For more details about the model, see Learn more about the model.

Registration is required to use MedImageInsight Premium. Access will be granted according to Microsoft’s eligibility criteria. To request access, submit this form.

Prerequisites

• An Azure subscription with access to Microsoft Foundry. If you don’t have one, create one for free.
• A Foundry project. If you don’t have one, create a project.
• Permission to view models and create or use deployments in your project.
• Access to deploy MedImageInsight Premium in the model catalog for your project. To request access, submit this form.
• Required role-based access control for model deployment and endpoint use. For details, see Role-based access control in Foundry portal.
• A test image or small test set for evaluation.
• A client for test calls, such as REST tooling or an SDK-capable app environment.

Deploy MedImageInsight Premium in Foundry

Deploy the model from the Foundry model catalog so that you can invoke it from your application or test client.

1. Sign in to Microsoft Foundry. Make sure the New Foundry toggle is on. These steps refer to Foundry (new).

1. Select your subscription and Foundry resource.
2. Select Discover > Models.
3. Search for MedImageInsight Premium and open the model card.
4. Select Deploy.
5. Review the available terms and deployment settings in your tenant.
6. Enter a deployment name and create the deployment.
7. Wait for deployment status to show Succeeded.
8. Copy the endpoint URL, deployment identifier, and authentication settings.

If deployment fails, check Known issues in Microsoft Foundry for current limitations and workarounds. Common causes include missing role assignments, region mismatch, and offer access that isn’t enabled for your tenant.

Test the deployment

After deployment succeeds, you can validate the endpoint by sending test requests with text or image inputs to generate embeddings.

Sample request payload

Send a POST to /providers/microsoft/v2/embed on your deployment URL. Use the texts field for text inputs or the images field for images encoded as base64 data URIs.

REST

    POST https://<your-endpoint>/providers/microsoft/v2/embed
    Authorization: Bearer <your-api-key>
    Content-Type: application/json

    {
      "model": "MedImageInsight-Premium",
      "texts": ["x-ray chest anteroposterior Atelectasis."]
    }

Reference for REST API

The following API reference shows the request payload, headers, response schema, and error codes for the MedImageInsight Premium embedding endpoint. Use it to send authenticated requests, confirm your deployment is reachable, and as a reference to build applications using the API.

Request headers

• Authorization: Bearer <your-api-key>
• Content-Type: application/json

Request body

Field	Type	Required	Default	Description
model	string	Yes	—	Served model name for the deployment.
texts	string array	Conditional	—	Text inputs to embed. Provide texts or images.
images	string array	Conditional	—	Images as base64 data URIs (for example, data:image/png;base64,...). Provide texts or images.
embedding_types	string array	No	["float"]	Accepted values: float (recommended), binary, ubinary, base64. See Embedding formats.

• One modality per request. Provide texts or images, not both. If both fields are present, the endpoint silently processes one and ignores the other.
• Batching is supported. You can send multiple items in a single call. The response returns one vector per item, ordered to match the input array.
• Supported image formats: PNG and JPEG.
• Image preparation. If your source data is DICOM, convert it before sending. The model expects single 2D images; for volumetric studies (CT, MRI), select the relevant slice and apply standard windowing for the modality before encoding. For reference DICOM-to-image conversion and modality-specific preprocessing utilities, see the Healthcare AI Examples repository.

Embedding formats

The response key matches the requested type (for example, embeddings.binary when you request binary).

• float (recommended): list of 1024 float values per vector. The canonical embedding.
• base64: the float vector encoded as raw float32 bytes, base64-encoded. Lossless equivalent of float, smaller wire payload.
• binary / ubinary: 128 bytes per vector — the 1024 sign bits of the float vector, packed 8 per byte. 32× smaller but lossy (only preserves sign). Suitable for fast approximate similarity search using Hamming distance.

Minimal text-only request

{
  "model": "MedImageInsight-Premium",
  "texts": ["x-ray chest anteroposterior Atelectasis."]
}

Full image-embedding request

{
  "model": "MedImageInsight-Premium",
  "images": ["data:image/png;base64,<base64-encoded-image>"]
}

Response body

Field	Type	Description
embeddings.float	array of float arrays	List of embedding vectors. Each vector is a list of 1024 floating-point numbers. The outer list has one vector per input item, ordered to match your texts or images.
id	string	Unique identifier for the request.
texts	string array	Echo of the input texts, if provided.
meta	object	Request metadata including API version and billed units.
response_type	string	Response format identifier.

{
  "id": "...",
  "embeddings": {
    "float": [[0.0123, -0.0456, "... (1024 floats total)"]]
  },
  "texts": ["x-ray chest anteroposterior Atelectasis."],
  "meta": { ... },
  "response_type": "embeddings_by_type"
}

Response codes

Status	Condition
200	Request succeeded.
400	Missing or invalid request field. Confirm that model is present, images values use the data:image/...;base64, prefix, and any provided embedding_types values are valid.
401 or 403	Confirm that you’re using a key for the same resource as the endpoint.
404	Confirm that the endpoint starts with https:// and matches the endpoint shown for your deployment.

Learn more about the model

MedImageInsight Premium (preview) generates rich, semantically meaningful embeddings of medical images across nine imaging modalities, including X-ray, CT, MRI, ultrasound, dermatology, ophthalmology, pathology, mammography. These embeddings power downstream workflows: similarity search, classification, outlier detection, drift monitoring, dataset curation, and multimodal retrieval-augmented generation. Outputs are intermediate signals that feed into a customer-built application; they are never a clinical determination on their own.

The premium model is a closed-weight, serverless offering with expanded capabilities and improved performance. For more information about differences between legacy and Premium models, see Legacy and Premium healthcare models.

For license, transparency, and intended-use details, see the MedImageInsight Premium model card.

Common use cases

Each of the following use cases assumes qualified human review as part of the workflow before any approval or action occurs

• Image similarity search across hospital PACS archives
• Dataset curation and triage for AI/ML pipelines
• Outlier detection and study-level QA
• Drift monitoring for deployed imaging models
• Embedding-based classification for narrow downstream tasks (fracture detection, lesion characterization, modality routing)

For runnable code examples covering zero-shot classification, adapter training, outlier detection, and more, see the Healthcare AI Examples repository on GitHub.

Review safety requirements

MedImageInsight Premium (preview) generates embeddings for medical images and text in a shared latent space, suitable for similarity search, retrieval, and downstream classification workflows.

MedImageInsight Premium is a model service, not a standalone clinical application. It’s intended for organizations and developers building healthcare imaging solutions, including healthcare providers, independent software vendors, systems integrators, partners, enterprise AI teams, and data science teams. The service is hosted and accessed by authenticated endpoints; customers don’t receive raw model weights.

Embedding outputs can support powerful workflows, but they don’t provide definitive clinical truth. Model performance can vary based on data representativeness, modality, acquisition parameters, workflow design, site-specific practice, prompt wording, preprocessing, image quality, and distribution shift.

Before implementation, define your workflow controls:

1. Validate that the model is fit for your intended use, modalities, and data sources.
2. Keep appropriate human governance and review in workflows where outputs inform decisions.
3. Evaluate performance on representative data before use and after material workflow changes.
4. Monitor data drift, out-of-distribution behavior, quality, fairness, and operational performance over time.
5. Confirm privacy, security, retention, logging, and access controls for sensitive healthcare data.

Use MedImageInsight Premium only in assistive workflows with appropriate human oversight. It isn’t intended for:

• Out-of-the-box clinical use
• Autonomous clinical decision-making.
• Use without qualified human oversight where outputs inform health or medical decisions.
• Treating similarity scores, embeddings, classifications, or search results as definitive clinical truth without validation.
• Use cases that require guarantees of perfect accuracy, completeness, fairness, or stability across all populations, devices, sites, or imaging protocols.
• Emergency, triage, or time-critical workflows unless your organization has independently validated the complete workflow and implemented appropriate controls.
• Any workflow where model output could be acted on without mitigation for performance variability, drift, or data quality issues.

Data, privacy, and security considerations

MedImageInsight Premium might be used in workflows that involve sensitive healthcare data, including medical images and associated text. You are responsible for configuring and operating your applications to meet privacy, security, compliance, and data governance obligations.

Use of MedImageInsight Premium is subject to the preview license and might also be subject to other terms and conditions. For licensing information, see the MedImageInsight Premium model card.

Related content

• Healthcare AI examples (GitHub)
• How to use MedImageInsight healthcare AI model for medical image embedding generation (classic)
• Customize a premium healthcare AI model with fine-tuning
• Model catalog and collections in Foundry portal
• Authentication and authorization options in Foundry
• Integrate Microsoft Foundry with your applications