Announcing the Beta Release of XNAT Machine Learning

[Will Horton]

Announcing the XNAT Machine Learning Suite

For the last several months, we have been working on adding support for Machine Learning and model training workflows to XNAT, in collaboration with NVIDIA, Radiologics, and the ICR Imaging Informatics group. We first demonstrated a proof of concept using models and APIs from the NVIDIA Clara™ Imaging framework with accelerated GPU computing at the 2019 RSNA conference, and have since been working to standardize these features and also add functionality.

What Can You Do With the XNAT Machine Learning Suite?
With XNAT ML (Beta), you can assemble training-specific collections of your imaging data files in a given project, draw new segmentations and annotations on that data, install and configure a training model, then train that model on your imaging data to enable an AI-assisted annotation workflow. Future releases in the XNAT ML line will enable model sharing and inference.

The XNAT ML (Beta) distribution is built on a pre-release version of XNAT 1.8, which is built on Java 8, PostgreSQL 12, and includes a number of enhancements, plugins, and custom components to support the model training workflow. All of these components are wrapped up in a docker-compose package, to give you a single-step installation process. This package includes documented workflows with known limitations. We are looking for your feedback on how we can improve this toward a full production release.

Download XNAT ML (Beta)

View the XNAT Machine Learning Documentation

[Shenjun Zhong]

We were internally testing the XNAT-Ml Suite, just wondering what is the Dataset Criteria Definitions?, what should be input to the Criteria (File Matchers Only)? An example of that?

Cheers

[Rick Herrick]

Hi Zhong,

Here's the documentation for creating a dataset: https://wiki.xnat.org/display/ML/Defining+Parameters+for+your+Dataset

One thing I'd emphasize is at the very bottom of the page, in the paragraph starting "The TaggedResourceMap lets you specify resource tags..."

Specifically, the value specified for SeriesDescription looks at 3 separate attributes on the XNAT image scan data type, type, series_description, and series_class. That means looking for a value like "T1 flair" amounts to a query like "type = 'T1 flair' OR series_description = 'T1 flair' OR series_class = 'T1 flair'".

The other important thing mentioned there is that the search value itself specifies how it is to be searched for:

• A value that contains the '%' character is searched with an SQL LIKE, e.g. 'T1%' becomes "xxx LIKE 'T1%'"
• A value that starts and ends with '/' is treated as a case-sensitive regular expression, e.g. '/T1.*/' becomes "xxx ~ 'T1.*'"
• A value that starts with '/' and ends with '/i' is treated as a case-insensitive regular expression, e.g. '/T1.*/' becomes "xxx ~* 'T1.*'"
  
• Any other value is searched with a standard comparison, e.g. 'T1 flair' becomes 'xxx = 'T1 flair'"

The idea is that you can select data for your dataset by series description and various attributes on resources associated with series. Try starting with the JSON in the Dataset Definition Criteria Payload example, adapt it to your data, and let us know how it works for you. This is very much a first pass at this functionality, so we're interested in how well it works (or how not well it works, as the case may be) and what other functionality people need to define and compile their datasets for model training and other uses.

[Shenjun Zhong]

thanks, Rick,

We are looking into that. Hopefully we can provide some early stage user feedback regarding this feature.

Cheers

--
You received this message because you are subscribed to the Google Groups "xnat_discussion" group.
To unsubscribe from this group and stop receiving emails from it, send an email to xnat_discussi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/xnat_discussion/4bbf8241-8bbe-41bc-a049-ede55fd6f42cn%40googlegroups.com.

--

Shenjun Zhong, PhD

Research Scientist

MASSIVE

Monash Biomedical Imaging (MBI)

Faculty of Information Technology (FIT)

Monash University, Australia