AIR Datasets and Root Results - Proposal

1. Proposed Workitem: AIR Datasets and Root Results

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

2. The Problem

The AIR Profile specifies predictable encoding of image analysis results for reliable receipt, parsing and display by consumers (Image Displays).

First, while the AIR Profile includes some text examples at the end of Annex A, it can be challenging for implementers (per comments from Lynn Felhofer and Herman Oosterwijk) to understand how the specification applies in their case and to correctly create conformant objects without a fuller set of examples as digital datasets.

Second, broad adoption of image analysis AI poses the next problem for Image Displays, which is how to navigate large result datasets:

• Expanded use of AI algorithms may produce very large collections of results for a given study
• Image Displays need to present that information to radiologists/clinicians
• Large result sets have logical structure/hierarchy that would help clinicians navigate and review the data

• Where to start; “Root”/“Summary” findings
• “Summary” findings are supported/derived from sub-sets of “sub-findings” (“drill-down”, “next layer”)

• The open-ended nature of results being provided for display mean image displays will be hard pressed to organize it themselves
• Without such organization, navigating large result sets will be labor intensive for radiologists, leading them to either waste time or ignore information.

So:

• The Image Display needs a simple way to access & leverage the hierarchy/structure
• The Creator likely knows that structure but needs a way to communicate it

• Advanced “data organizing” software could also create summaries and structure for results from multiple algorithms

3. Key Use Case

CT Lung Screening Example:

• An AI analysis package detects 8 nodules
• For each detected nodule,

• a segmentation algorithm generates a segmentation and a centroid location
• a third algorithm estimates the size, the solidity, the margin, and the LungRADS assessment

• it also generates an overall LungRADS™ score.
• another algorithm (outside the LungRADS package) generates a result that pneumonia is present

• and stores an associated saliency map

These roughly 40+ results are stored to the study (cardiac calcification and other screenings were not run on this study).

Strawman concept:

• The Lung Screening package stores a Root Result object with a root finding of (LungRADS = Category 3)

• the root result references the "layer 2" findings (the 8 nodule locations and their LungRADS values)
• the root result references the "layer 3" findings for each nodule (segmentation, and assessments of the size, solidity, and margin)

• The Pneumonia application stores a Root Result object with a summary finding of (“pneumonia present”)

• that root result references the saliency map instance.

• The Image Display identifies two Root Result instances in the study and presents the layer 1 findings in the initial overlay

• LungRADS = Category 3
• Pneumonia present

• The radiologist may want to gain confidence in a layer 1 "summary finding", or to comprehend more details/nuances of the finding.
• The radiologist selects a layer 1 finding (LungRADS = Category 3) and its layer 2 findings are presented

• 8 nodule locations annotated with individual LungRADS scores

• The radiologist selects a layer 2 finding (one of the nodules) and it's layer 3 findings are presented

• a nodule segmentation
• the nodule size, solidity, and margin assessment

Might be interesting for the results to include a confidence and "potential significance" to assist in filtering/layering/organizing/prioritizing/progressive disclosure. Other navigation paradigms can be discussed. Some Display behaviors would likely be customizable to suit radiologist preferences. Ideally, some displays will develop much more sophisticated analysis and logic, or more advanced configurations, and more advanced navigation and display, while the Root Results provide a first simple step up from the flat list of findings.

The goal is to facilitate some basically useful navigation without Displays having to be customized for each AI algorithm. Similar to the use of primitives.

4. Standards and Systems

DICOM SR analogous to Key Object Selection seems like a valid approach.

5. Discussion

This problem was identified during development of the AIR Profile and a “Root result” object was proposed.

• Some public comments supported the need and value
• Others comments challenged that the proposed mechanism had not been fully thought through or the potential complexities mapped out

Some form of DICOM SR object seemed like a valid approach, but TODO

One piece of proposed work will be to create example datasets. (pilot the proposed improvement from the last retrospective)

A second piece of proposed work will be to assess how to make dataset organization tractable for relatively simple Image Displays.

An organizing object will be evaluated. If it seems workable, creating and using said object could be added as a Named Option in the AIR Profile for the Evidence Creator and Image Display actors.

It would not make sense to mandate that all the results in the study always be cataloged in a single result object index, since that would require continually revising the catalog object each time one of many algorithms stored new results to the study. It would be challenging, not to mention handling competing updates by when multiple algorithms happen to complete at the same time.