How to use MedImageInsight healthcare AI model for medical image embedding generation

In this article, you learn how to deploy MedImageInsight from the model catalog as an online endpoint for real-time inference. You also learn how to issue a basic call to the API. The steps you take are:

• Deploy the model to a self-hosted managed compute.
• Grant permissions to the endpoint.
• Send test data to the model, receive, and interpret results.

MedImageInsight - the medical imaging embedding model

MedImageInsight foundation model for health is a powerful model that can process a wide variety of medical images. These images include X-Ray, CT, MRI, clinical photography, dermoscopy, histopathology, ultrasound, and mammography images. Rigorous evaluations demonstrate MedImageInsight's ability to achieve state-of-the-art (SOTA) or human expert-level performance across classification, image-to-image search, and fine-tuning tasks. Specifically, on public datasets, MedImageInsight achieves or exceeds SOTA performance in chest X-ray disease classification and search, dermatology classification and search, Optical coherence tomography (OCT) classification and search, and 3D medical image retrieval. The model also achieves near-SOTA performance for histopathology classification and search.

An embedding model can serve as the basis of many different solutions—from classification to more complex scenarios like group matching or outlier detection. The following animation shows an embedding model being used for image similarity search and to detect images that are outliers.

Prerequisites

• An Azure subscription with a valid payment method. Free or trial Azure subscriptions don't work. If you don't have an Azure subscription, create a paid Azure account to begin.

• If you don't have one, create a hub-based project.

• Azure role-based access controls (Azure RBAC) to grant access to operations in Azure AI Foundry portal. To perform the steps in this article, your user account must be assigned the Azure AI Developer role on the resource group. For more information on permissions, see Role-based access control in Azure AI Foundry portal.

Deploy the model to a managed compute

Deploying to a self-hosted managed inference solution lets you customize and control all the details about how the model is served. You can deploy the model from its model card in the catalog UI of Azure AI Foundry, Azure Machine Learning studio, or deploy it programmatically.

To deploy the model through the UI:

1. Go to the model catalog.

2. Search for the MedImageInsight model and select its model card.

3. On the model's overview page, select Use this model to open the deployment window.

   Note

   For deployment to a self-hosted managed compute, you must have enough quota in your subscription. If you don't have enough quota available, you can use our temporary quota access by selecting the option I want to use shared quota and I acknowledge that this endpoint will be deleted in 168 hours.

4. Select the checkbox in the deployment window to use the temporary shared quota.

5. The deployment window displays settings that include a virtual machine selection, an endpoint name, and a deployment name. You can modify these settings. Once you're satisfied with them, select Deploy.

To deploy the model programmatically, see How to deploy and inference a managed compute deployment with code.

Work with an embedding model

In this section, you consume the model and make basic calls to it.

Use REST API to consume the model

Consume the MedImageInsight embedding model as a REST API, using simple GET requests or by creating a client as follows:

from azure.ai.ml import MLClient
from azure.identity import DefaultAzureCredential

credential = DefaultAzureCredential()

ml_client_workspace = MLClient.from_config(credential)

In the deployment configuration, you choose an authentication method. This example uses Azure Machine Learning token-based authentication. For more authentication options, see the corresponding documentation page. Also, the client is created from a configuration file that's created automatically for Azure Machine Learning virtual machines (VMs). Learn more on the corresponding API documentation page.

Make basic calls to the model

After deploying the model, use the following code to send data and get embeddings.

import base64
import json
import os

endpoint_name = "medimageinsight"
deployment_name = "medimageinsight-v1"

sample_image_xray = os.path.join(image_path)

def read_image(image_path):
    with open(image_path, "rb") as f:
        return f.read()

data = {
    "input_data": {
        "columns": ["image", "text"],
        #  IMPORTANT: Modify the index as needed
        "index": [0],
        "data": [
            [
                base64.encodebytes(read_image(sample_image_xray)).decode("utf-8"),
                "x-ray chest anteroposterior Pneumonia",
            ]
        ],
    },
    "params": {"get_scaling_factor": True},
}

# Create request json
request_file_name = "sample_request_data.json"
with open(request_file_name, "w") as request_file:
    json.dump(data, request_file)

response = ml_client_workspace.online_endpoints.invoke(
    endpoint_name=endpoint_name,
    deployment_name=deployment_name,
    request_file=request_file_name,
)

Use MedImageInsight REST API

The MedImageInsight model works with a simple single-turn interaction. One request gets one response.

Request schema

The request payload is a JSON-formatted string that contains the following parameters:

The input_data object contains the following fields:

The params object contains the following fields:

Request example

A simple inference requesting embedding of a single string

{
  "input_data": {
    "columns": [
      "image",
      "text"
    ],
    "index":[0],
    "data": [
      ["", "a quick brown fox jumps over the lazy dog"]
    ]
  },
  "params": {}
}

Request for embedding of an image and a string, requesting for return of the scaling factor

{
    "input_data": {
        "columns": ["image", "text"],
        "index": [0],
        "data": [
            [
                "iVBORw0KGgoAAAANSUhEUgAAAAIAAAACCAYAAABytg0kAAAAAXNSR0IArs4c6QAAAARnQU1BAACx\njwv8YQUAAAAJcEhZcwAAFiUAABYlAUlSJPAAAAAbSURBVBhXY/gUoPS/fhfDfwaGJe///9/J8B8A\nVGwJ5VDvPeYAAAAASUVORK5CYII=\n",
                "x-ray chest anteroposterior Pleural Effusion"
            ]
        ]
    },
    "params": {
        "get_scaling_factor": "true"
    }
}

Response schema

The response payload is a JSON-formatted string that contains the following fields:

Response example

A simple inference requesting embedding of a single string

{
  "image_features": [[0.029661938548088074, -0.027228673920035362, ... , -0.0328846238553524]],
  "text_features": [[0.0028937323950231075, 0.004354152828454971, -0.0227945726364851, ..., 0.002080598147585988]],
  "scaling_factor": 4.516357
}

Other implementation considerations

The maximum number of tokens processed in the input string is 77. The system removes any tokens beyond 77 before passing the input to the model. The model uses a Contrastive Language-Image Pre-Training (CLIP) tokenizer, which uses about three Latin characters per token.

The model embeds the submitted text into the same latent space as the image. As a result, strings describing medical images of certain body parts obtained with certain imaging modalities are embedded close to such images. Also, when building systems on top of a MedImageInsight model, make sure that all your embedding strings are consistent with one another (word order and punctuation). For best results with the base model, strings should follow the pattern <image modality> <anatomy> <exam parameters> <condition/pathology>., for example: x-ray chest anteroposterior Atelectasis..

If you fine-tune the model, you can change these parameters to better suit your application needs.

Supported image formats

The deployed model API supports images encoded in PNG format.

When the model receives the images, it preprocesses the images by compressing and resizing them to 512x512 pixels.

The preferred compression format is lossless PNG, containing either an 8-bit monochromatic or RGB image. For optimization purposes, you can perform resizing on the client side to reduce network traffic.

Learn more from samples

MedImageInsight is a versatile model that you can apply to a wide range of tasks and imaging modalities. For more specific examples of solving various tasks with MedImageInsight, see the following interactive Python notebooks.

Getting started

• Deploying and Using MedImageInsight: Learn how to deploy the MedImageInsight model programmatically and issue an API call to it.

Classification techniques

• Building a Zero-Shot Classifier: Discover how to use MedImageInsight to create a classifier without the need for training or large amount of labeled ground truth data.

• Enhancing Classification with Adapter Networks: Improve classification performance by building a small adapter network on top of MedImageInsight.

Advanced applications

• Inferring MRI Acquisition Parameters from Pixel Data: Understand how to extract MRI exam acquisition parameters directly from imaging data.

• Scalable MedImageInsight Endpoint Usage: Learn how to generate embeddings of medical images at scale using the MedImageInsight API while handling potential network issues gracefully.

Related content

• MedImageParse models for medical image segmentation
• CXRReportGen for grounded report generation