AI workflows
An AI workflow ties an AI service to the source its results come back as, and tracks every study sent until its results arrive. Configuration › AI workflows.
Setting one up
Section titled “Setting one up”- The AI service is a destination (DICOM or DICOMweb). Route studies to it like to any destination, best with a route that waits for the whole study.
- To send it de-identified studies, give that destination a de-identification profile.
- Its results arrive as a source: one matching the AI service as it sends back (its AE title and address, or its DICOMweb sender). Give that source a route to where results should go: the PACS, say.
- Add the workflow: the AI service, the results source, how long results may take (expected within), and how long after the last result a study counts as complete.
Every study is a job
Section titled “Every study is a job”History › AI jobs lists every study sent, across all nodes:
| State | Means |
|---|---|
| Sending | Instances still queued for the AI service. |
| Waiting for results | Sent; no result yet. |
| Results coming in | At least one result arrived. |
| Complete | No further result for the workflow’s quiet time. |
| Timed out | No result within expected within: alerted. A late result still moves it on. |
| Failed | Nothing could be delivered to the AI service. |
Jobs are kept centrally, so results may come back to any node.
Results get their identity back
Section titled “Results get their identity back”When the AI service gets de-identified studies, each node records the identity and the UIDs it sent. Results (new instances in the study, or edited copies) are given the patient’s identity and the original UIDs back, references in sequences included, before the results source’s routes send them on. The PACS files them with the original study.
A result for a study not known yet is held on the node and tried again; after an hour it is refused (and quarantined), so nothing ever goes on under a pseudonym. Without de-identification, results go on as they come.