Setting up Model Proxy

December 30, 2021

TLDR:

If you didn't get your TSL certificate file execute Generate local TSL first.

git clone https://github.com/zhiva-ai/model-proxy-example.git

docker-compose up

Requirements

Before we start please make sure your server has access to:

Get the server code

You can either clone this repo

git clone https://github.com/zhiva-ai/model-proxy-example.git

or download it directly from zhiva.ai PACS server.

Local PACS and Model API

To use model proxy you have to get something that proxy could use. There are two main objects handled by proxy server.

The first one is Model API (setup here). Model API is responsible for inferences, and it is the whole reason the proxy exists. You have to define your model APIs as described in models.json section.

The second one is PACS Server (setup here). Server is required to extract DICOM data for the Model APIs. It has to be server that supports DICOMWeb standard. You have to define your server as described in servers.json section.

Setup your local server

Before you start make sure you Generate local TSL.

At this point you should have 1 .crt files and 1 .key file. Check this by calling

 ls *.{crt,key}

from your main directory. It should return following result zhiva.crt zhiva.key.

Build the server

docker-compose up

You proxy server should be available at https://localhost:8002.

Access from within internal network

If you have more than one computer inside your network (or VPN connection), then you can share the server settings with them. To check the server address please run the following command:

Linux or Mac:

ifconfig

Windows

ipconfig

and look for the setting with the inet value that starts with 192.168.. That should by your address in the local network. You should be able to access the server from 192.168.x.x:8002.

Requesting inference for given model

To get an inference using selected Model API you should follow Inference API Docs. This example specifies the request for series segmentation. Model UUID is the same model as defined in models.json.

If you have mode than one server defined in ./servers.json then you can select which one should provide DICOM data by setting ?server query parameter like:

/segmentations/{model-uid}/studies/{study-uid}/series/{series-uid}?server=63140137-a63d-4ad1-b489-3479bc43387c

Value of this parameter is the UUID key assigned to the server inside servers.json.

E.g. To get inference (segmentation) for Study 1.3.6.1.4.1.25403.345050719074.3824.20170125095438.5 and Series 1.3.6.1.4.1.25403.345050719074.3824.20170125095449.8 using Model API with UUID cf8063dc-d4bb-4b68-8e74-eba77c248c8b and PACS server with UUID b5208f48-f748-4339-9402-d1148dab571e the request will look like:

https://localhost:8011/segmentations/cf8063dc-d4bb-4b68-8e74-eba77c248c8b/studies/1.3.6.1.4.1.25403.345050719074.3824.20170125095438.5/series/1.3.6.1.4.1.25403.345050719074.3824.20170125095449.8?server=b5208f48-f748-4339-9402-d1148dab571e

to get the same inference but using default PACS server just remove server query parameter:

https://localhost:8011/segmentations/cf8063dc-d4bb-4b68-8e74-eba77c248c8b/studies/1.3.6.1.4.1.25403.345050719074.3824.20170125095438.5/series/1.3.6.1.4.1.25403.345050719074.3824.20170125095449.8

servers.json (add DICOMweb server)

path: ./servers.json

Keeps all definitions of available pacs servers. It has to be a valid dicomweb server. Each sever has to use uuidv4 which will be used to access it. You can generate such an uuid from online generator.

{
  [{uuidv4}]: {
    "uri": string ("https://valid.server.uri/to/dicomweb"),
    "isDefault": boolean (default false),
    "statusUri": string ("https://valid.server.uri/to/status/page"),
    "authType": string ("" | "basicAuth"),
    "login": string,
    "password": string,
    "token": string,
    "signInServer": string ("https://localhost/auth/sign-user"),
    "signInTokenServer": string ("https://localhost/auth/generate-access-token"),
    "refreshServer": string ("https://localhost/auth/refresh-token"),
  }
}
  • uri - URI of dicomweb server
  • isDefault - boolean which indicates the default server used when request doesn't specify the uuidv4 value. One server should have this set to true.
  • statusUri - (optional)URI to status page. This can be any page accessible through GET request that returns stats 200 if server is working correctly. If you don't have any specific page, you can just point to list of studies: https://valid.server.uri/to/dicomweb/studies.
  • authType - Authentication method ("" | "basicAuth" | token | zhivaAuth)
  • login - (optional) Login to use when authType or zhivaAuth is set
  • password - (optional) Password to use when authType or zhivaAuth is set
  • authToken - (optional) token to use when token or zhivaAuth is set
  • signInServer - (optional) zhiva sign in server to use when zhivaAuth is set and login with password are provided (usually https://localhost/auth/sign-user)
  • signInTokenServer - (optional) token sign in server to use when token or zhivaAuth is set (https://localhost/auth/generate-access-token for Local PACS with JWT)
  • refreshServer - (optional) refresh token server to use when token or zhivaAuth is set (https://localhost/auth/refresh-token for Local PACS with JWT)

models.json (add model API)

path: ./models.json

Stores all definitions of inference models. Each model might be available for different type of task. Each model has to use uuidv4 which will be used to access it. You can generate such an uuid from online generator.

{
  [{uuidv4}]: {
    "uri": string ("https://valid.model.uri/to/inference"),
    "task": string ("segmentation" | "annotation" | "prediction"),
    "supports": Array<string> (list of supported paths, eg. ["/studies/series"]),
    "statusUri": (optional) string ("https://valid.model.uri/to/status/page")
  }
}
  • uri - URI of the inference model API (eg. http://localhost:8011/predict)
  • task - Task which model performs (one of segmentation | annotation | prediction)
  • supports - List of available paths that model supports:
/studies
/studies/series
/studies/series/instances
  • statusUri - URI to status page. This can be any page accessible through GET request that returns stats 200 if model is working correctly.

Status check

After defining your models and servers you can check if all of them are available through proxy. This is possible by accessing https://localhost:8002/status page. This page is only visible when your docker-compose.yml configuration has ZHIVA_SHOW_STATUS_PAGE variable set to true:

    environment:
      - ZHIVA_SHOW_STATUS_PAGE=true

Authentication

Basic auth

If your PACS server requires an authentication please provide correct credentials (login and password) as well as change the authType value. If you're using our Local Server you can read more about securing your server in the Server Authentication section.

Example setting:

  "b1407f18-575d-487c-bc0c-640c6da651bc": {
    "uri": "https://localhost/zhiva/pacs",
    "statusUri": "https://localhost/zhiva/pacs/studies",
    "authType": "basicAuth",
    "login": "zhiva",
    "password": "35&Q39Nj&i@Eyk6P"
  }

This setting with work with authentication defined in our local PACS server configuration. It uses login and password to generate authentication token which is then used by the server to access data from PACS.

Zhiva auth

If you're using our Local PACS with JWT you should use our standard authentication method. First, you have to generate User account without "Can edit PACS resources?" access (unless proxy has to be able to edit data on PACS). Then just enter this user's credentials to your server config like:

  "b1407f18-575d-487c-bc0c-640c6da651bc": {
    "uri": "https://localhost/zhiva/pacs",
    "statusUri": "https://localhost/zhiva/pacs/studies",
    "authType": "zhivaAuth",
    "login": "your_username",
    "password": "your_secret_password",
    "signInServer": "https://localhost/auth/sign-user",
    "signInTokenServer": "https://localhost/auth/generate-access-token",
    "refreshServer": "https://localhost/auth/refresh-token"
  }

And you're good to go! Now, proxy server will use provided user to generate token and store it inside Docker's persistent volume.

External Token

If you have your own token provider you can use it with one-off token authentication. You have to generate Auth Token and enter it into the server's config.

  "b1407f18-575d-487c-bc0c-640c6da651bc": {
    "uri": "https://localhost/zhiva/pacs",
    "statusUri": "https://localhost/zhiva/pacs/studies",
    "authType": "token",
    "authToken": "YOUR_AUTH_TOKEN",
    "signInTokenServer": "https://localhost/auth/generate-access-token",
    "refreshServer": "https://localhost/auth/refresh-token" //optional
  }

This authToken is going to be sent to signInTokenServer which should return Access Token (in Authentication header) and optional Refresh Token in response's body (Content-Type: text/html). Access Token should contain following fields:

{
  "name": string,
  "username": string,
  "canEditResource": boolean,
  "iat": number,
  "exp": number,
  "iss": string
}

When Auth Token expires then (if provided) Refresh Token will be used to retrieve new token from refreshServer.