AI Results - Brief Proposal

1. Proposed Workitem: AI Results (AIR) Profile

• Proposal Editor: Kevin O'Donnell
• Editor:
• Domain: Radiology

2. The Problem

There is significant interest and activity in the application of AI methods (deep learning, etc) to the analysis of medical images.

The results of such AI analysis need to be presented to radiologists in the context of reading the corresponding study. This depends on a degree of integration between AI result generation products and radiology reading workstation systems/software.

Additional complexity comes from the fact that radiologists interpreting studies likely expect AI-generated results to supplement rather than replace more traditional image analysis results and thus a given study will be composed of the acquired images, the AI results, and traditional clinical data. In a similar vein, many AI results may be results that a human would otherwise have produced, and further those human results may be used as training data for the AI, and AI results may be used by other AIs in an adversarial network. So some convergence of result encoding would be helpful.

For AI to productively live up to it's promise, the datasets must be reliably, and conveniently, assembled and managed.

3. Key Use Case

The goal is for AI packages to be able to store AI results that can be retrieved and presented consistently by a variety of imaging display systems.

The display would present the result in the context of the medical imaging study to which it applies.

To minimize the implementation complexity for the displays, and to avoid having to change software for each new AI result, it would make sense to compose AI results from a reasonable set of primitives.

4. Standards and Systems

Result Creators - AI algorithms running in PACS, cloud, processing servers, standalone workstations, etc.

Storage - PACS, VNA, etc.

Result Displays - reading workstations (could argue this is a special case of consumer, but might want to have specific required display features/behaviors)

Result Consumers - databases, clinical analysis, report creators (import results)

Standards

DICOM SR - TID 1500

• TID 1500
• WG-23 is working to catalog result types and identify any gaps in TID 1500
• An initial set of primitives have been identified
• WG-23 AI Results Workplan Doc

DICOM Segmentations DICOM Sup XXX Simplified SR in JSON for AI

• David Clunie has a significant draft that will be presented at WG-06 in September. He has also been working on tooling to test/confirm the transcoding logic.

RSNA CDE (Common Data Elements)

• helps with longitudinal data and automation logic for reporting and decision support
• http://radelement.org

IHE Evidence Documents Profile

• Should review to see if there anything we should borrow/revive

5. Discussion

This fits the IHE model of profiles as tools for convergence.

• Display products that support those primitives can declare they are “AI-Ready”
• Users can motivate display vendors to be AI-Ready
• AI products that output their result using those primitives will know that a variety of displays will be able to present their results.
• Users can motivate AI vendors to conform since that will simplify deployment for the users

IHE Rad may want to consider defining some baseline presentation requirements or conventions for certain results.

Risks

• The JSON SR supplement could get hung up on technical hitches.