Clarus — a ~2 MB DICOMweb PACS + IHE AIW-I (UPS-RS) inference pipeline, with an open producer ABI

[clicker71]

We built a tiny DICOMweb origin server and a self-contained demo that runs a TotalSegmentator craniofacial task end-to-end over plain DICOMweb + IHE AIW-I — no proprietary glue, QIDO/WADO/STOW/UPS-RS only. We thought the MONAI Label team might find the integration surface useful.

The demo (Windows/Linux, ~2 MB; weights are not shipped — the first run downloads them):
https://github.com/clicker71/clarus-pacs/releases/tag/totalseg-demo-v0.1

What it does: STOW a head/jaw CT → the server mints an AIW-I workitem → a sidecar fetches the frames over WADO-RS, runs TotalSegmentator, and stores a native RTSTRUCT (MANDIBLE / TEETH_LOWER / TEETH_UPPER) back into the study. Run guide (EN/RU) is inside the pack.

The integration surface is ABI-README.md inside the pack. A "producer" is any program that:

• reads the workdir (raw pixel frames) and manifest.json (per-frame geometry, declared and validated against the source),
• writes a result.json,
  and the sidecar assembles conformant DICOM objects: SEG masks, PR overlays (GSPS), SR reports (narrative + TID 1500 measurement reports), RTSTRUCT contours, STL meshes.

Two ways MONAI Label could plug in:

1. Run against the server as a DICOMweb datastore: monailabel start_server --studies http://<clarus>/dicomweb.
2. Implement MONAI models as ABI producers — a small program reading the workdir/manifest and writing result.json.

Notes: the server and sidecar are closed-preview binaries for now (not open source yet), but the producer ABI is an open, documented contract — your integration is your code, under your license. Happy to answer any questions.

Best regards,
Clarus team — https://github.com/clicker71/clarus-pacs