AI White Paper Updates - Brief Proposal

1. Proposed Workitem: AI White Paper Updates

• Proposal Editor: Brad Genereaux
• Editor: TBD
• Contributors: Brad Genereaux, Neil Tenenholtz, Brian Bialecki, Chris Carr, Khaled Younis, Chris Roth, Kevin O’Donnell, Vikash Gupta, Barbaros S. Erdal
• Domain: Radiology

2. The Problem

Since the publication of the AI Interoperability in Imaging White Paper in October 2021, Template:Https://www.ihe.net/uploadedFiles/Documents/Radiology/IHE RAD White Paper AI Interoperability in Imaging.pdf, imaging AI technology has continued to evolve dramatically. These advancements have implications for a number of existing parts of the White Paper, and may call for some new sections to be added.

As with the original white paper, guidance of the interoperability community can help to understand and integrate these technologies, and efficiently connect these solutions at scale - for every AI use case, for every modality, for every disease, across every region of the world.

This proposal is focused on delivering incremental updates to the White Paper, rather than providing an addendum or a separate additional White Paper for implementers to review.

3. Key Use Case

Several advancements in the space of AI Imaging necessitate expanded or new sections within the AI White Paper, including:

• Generative Imaging AI: Generative imaging and synthetic images can be applied to development, administrative, and clinical workflows
  • Development: augmenting training data, …
  • Clinical: Synthetic contrast, synthetic aging, thin-slice generation, …
  • Administrative: Drift detection, de-identification, …
• Generative Text AI – the notion of calling upon Imaging AI applications from large language models (e.g. ChatGPT-style) prompts that leverage medical images and related metadata
  • How can generative text AI solutions call upon imaging inference solutions, to form orchestrated generative AI?
  • How does provenance work and relate back to source? How is output logged? Who is the “author” and how is that captured? How are human-based edits handled?
  • Handling unstructured model output as it relates to discrete structured standards and protocols - should output follow structured reports? How does tertiary formats get translated and do those need to be structured / stored?
• Multimodal AI - the ability of a model to simultaneously incorporate multiple types of inputs (images, narrative, data) is of particular relevance to medical imaging
• Foundational and Open Source/Open Access Models - the proliferation of base models and the decreased training/tuning costs expands the possibility of sites to develop custom/localized models for use. Are there aspects of the white paper that should address/discuss this more directly?
• Continuous Learning - the ability for AI models to be re-trained on the fly - what are the systems and interoperability needs to automate the feedback and re-training process?
• Cross-enterprise collaborative AI, including federated learning, federated dataset discovery, federated validation, federated inference, and federated prompt training.
  • Model and/or container sharing
• Image data conversion - DICOM to NIfTI, Numpy, etc, where are the issues and interoperability boundaries

These will be introduced in the relevant sections of the AI Interoperability in Imaging White Paper, and other sections of the White Paper will be amended as necessary by the editor team.

4. Standards and Systems

Systems involved:

• PACS, RIS, EMR, AI Orchestrator, AI Performer

Standards involved:

• DICOMweb - UPS-RS, WADO-RS, QIDO-RS, STOW-RS
• HL7 - ImagingStudy, Observation, DiagnosticReport

5. Discussion

1. The scope of updates goes beyond that of a CP; There is committee effort needed to assess and write text
2. Is an update to the current White Paper the right path forward, or should it be a standalone document?
   1. If this is an update, do we then need to review all sections for current relevance, etc.