Artificial Intelligence (AI) in medical imaging is important from an interoperability standpoint. Just as modalities are producing DICOM instances, pre-processing and post-processing systems are creating secondary objects, and humans are supplementing DICOM studies with their own objects, there is a role to play for machine learning and deep learning systems to interact within the medical imaging ecosystem. 

As with any other DICOM producer, however, it is important to meet a set of criteria to be good stewards within the ecosystem. Some important points to consider include:

  • On any DICOM producer website, including AI algorithm providers, there should be a published DICOM conformance statement outlining the types of objects it creates and how it interacts with them within the ecosystem
  • Consider important papers and presentations highlighting the following topics:
  • Ensure that objects bring created are conforming to specification, such as using the right SOP classes and appropriate metadata values. Examples include:
    • Secondary capture series, when derived from other series, may need to have the SOP Class Secondary Capture Image Storage (1.2.840.10008.5.1.4.1.1.7) used, rather than the original SOP class 
    • AI-generated objects likely should have the DICOM tag Image Type (0008,0008) with values for derived images of a secondary nature
    • Be sure to assign new unique identifiers to instances you create and be aware that it is generally not permitted to edit/overwrite DICOM instances through resubmitting altered objects with existing UIDs
  • You should consider contributing to DICOM WG-23 on Artificial Intelligence / Application Hosting, whose role is to identify or develop the DICOM mechanisms to support AI workflows, concentrating on the clinical context