# Idea - standard DICOMweb UPS-RS as the AI-task layer (trigger inference + read results in Slicer)

**URL:** <https://discourse.slicer.org/t/idea-standard-dicomweb-ups-rs-as-the-ai-task-layer-trigger-inference-read-results-in-slicer/48999>\
**Category:** Development\
**Tags:** ai\
**Created:** [November 20, 2026, 1:54pm UTC](https://discourse.slicer.org/t/idea-standard-dicomweb-ups-rs-as-the-ai-task-layer-trigger-inference-read-results-in-slicer/48999 "2026-11-20T13:54:37Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![clicker71](https://sea2.discourse-cdn.com/flex002/user_avatar/discourse.slicer.org/clicker71/32/83125_2.png) [@clicker71](https://discourse.slicer.org/u/clicker71)\
**Post date:** [November 20, 2026, 1:54pm UTC](https://discourse.slicer.org/t/idea-standard-dicomweb-ups-rs-as-the-ai-task-layer-trigger-inference-read-results-in-slicer/48999/1 "2026-11-20T13:54:37Z")

</div>

Hi everyone,  
I keep seeing excellent AI integrations land in Slicer as one-off extensions -  
the torchvulkan / GPU TotalSegmentator thread, MONAI-Label, per-model  
segmentation extensions. Each of them re-solves the same plumbing: run a model,  
get a result, show it in the scene. I would like to propose using the standard  
that already exists for exactly this plumbing - DICOMweb UPS-RS (PS3.18 section 11).  
The loop:  
1. The user selects data in Slicer and triggers a task.  
2. Slicer, as the Task Requester, creates a UPS workitem on the DICOMweb origin (POST /workitems), naming the study/series as the input.  
3. Any Task Performer - TotalSegmentator, MONAI, NVIDIA, clinfer, whatever -  
claims the workitem (Change State to IN PROGRESS), runs, writes its result  
back as DICOM SEG / SR / PR via STOW-RS, and marks it COMPLETED.  
4. Slicer subscribes (PS3.18 11.10/11.11, N-EVENT-REPORT over the event channel), sees COMPLETED, and loads the result over WADO-RS into the scene.  
No vendor-specific glue anywhere: the workitem is the contract. The same  
performer works from Weasis, from OHIF, from Slicer, unchanged.  
This is not hypothetical. The full loop already runs end-to-end against an open  
DICOMweb origin (Clarus) with an open performer (clinfer), driven from the  
Weasis viewer - screenshots in this Weasis PR comment:

> <https://github.com/nroduit/Weasis/pull/950#issuecomment-6219616602>
>
> Thanks for the explorer hooks - exactly the three extension points phase 2 needs… (context-menu contributor, thumbnail badges, DICOMweb origin).
> 
> Here is how we see phase 2 on top of them: Clarus as the origin, clinfer as the performer, one patient \`XRV-DEMO\` with a chest study (AI applies) and a knee study (it does not).
> 
> \*\*1. Availability.\*\* The badge is an availability indicator, not decoration: only the chest series carries it, because the catalogue task targets \`\*lung\*\` and the knee series has no matching BodyPartExamined. In the mergeable build this maps to \`AI\_TASK\` / \`AI\_RESULT\` through \`SeriesBadgeProvider\`.
> 
> !\[Badge on the chest series, none on the knee\](https://raw.githubusercontent.com/clicker71/clarus-pacs/main/images/weasis-demo-robot-visibility.png)
> 
> \*\*2. Done.\*\* Run (Create, RAD-80) -\> clinfer claims -\> COMPLETED. The badge flips to the result state, and "Show tasks" names the derived object by its own SOP Class UID (Secondary Capture), not a hard-coded one.
> 
> !\[Done with the result named\](https://raw.githubusercontent.com/clicker71/clarus-pacs/main/images/weasis-demo-task-done.png)
> 
> \*\*3. Cancel.\*\* While a task is queued or running, the menu offers Cancel (Request Cancellation, 11.8).
> 
> !\[Cancel while pending\](https://raw.githubusercontent.com/clicker71/clarus-pacs/main/images/weasis-demo-ability-to-cancel.png)
> 
> Two questions on the menu behaviour:
> 
> \- After COMPLETED, "Run" stays enabled (Create-if-none allows a re-run). Should the menu grey the entry once the task has a completed workitem, or keep it re-runnable? Today we disable it only while SCHEDULED / IN PROGRESS.
> \- The dialog lists each result as plain text. We would rather render it as a clickable entry that loads the object (image / SR / SEG) through the WADO-RS path - the performer already names the Series Instance UID (0020,000E) in the Output Information Sequence, so a direct retrieve URL is buildable. Is that the direction you want?
> 
> The workitem codes are minted under the private scheme \`99CLARUS\` (PS3.3 Table 8.8-1) - the deployment vocabulary, not part of Weasis.

For transparency, the same idea is filed for OHIF  
( [[Feature Request] UPS-RS User Agent (Task Requester + Watcher) to trigger AI workflows from the viewer · Issue #6369 · OHIF/Viewers · GitHub](https://github.com/OHIF/Viewers/issues/6369) ) and is under discussion with the dwv maintainer for a browser client.

Question for the Slicer community: would a UPS-RS User Agent extension for Slicer be useful - a worklist module that does the create / claim / subscribe / load loop, so any performer plugs in through one standard interface instead of each shipping its own workflow? Would the MONAI-Label / TotalSegmentator / NVIDIA AIAA folks see value in a common task contract?

I am happy to write the extension (Python, standard-only) and test it against Clarus and dcm4chee-arc as the origins, keeping it dependency-minimal.

Best regards,  
Daniil , [Clarus PACS](https://github.com/clicker71/clarus-pacs)

---

<div class="post-metadata">

**Author:** ![pieper](https://sea2.discourse-cdn.com/flex002/user_avatar/discourse.slicer.org/pieper/32/8_2.png) [@pieper](https://discourse.slicer.org/u/pieper)\
**Post date:** [November 21, 2026, 2:49pm UTC](https://discourse.slicer.org/t/idea-standard-dicomweb-ups-rs-as-the-ai-task-layer-trigger-inference-read-results-in-slicer/48999/2 "2026-11-21T14:49:57Z")

</div>

Yes, I think that would be a very welcome contribution. 👍
