Skip to content

RT research suite

RTpipeline and RTsegmentator cover the path from treatment-planning exports to analysis tables.

Tool Role
RTpipeline Organizes raw DICOM RT exports into validated patient courses and computes dose-volume, radiomics, robustness, multimodal and quality-control tables for whole cohorts.
RTsegmentator Runs segmentation models on single DICOM series and returns RTSTRUCTs checked against the source geometry, through a web interface or a versioned API. Documentation: kstawiski.github.io/rtsegmentator. 
 TPS export ──► RTpipeline organize ──► planning CT per course
                                             │
                                             ▼  upload
                                       RTsegmentator ──► RTSTRUCT + manifest
                                             │
                     verify checksums, series and ROI inventory
                                             ▼
               Segmentation_CustomModels/rtseg_<model>/rtstruct.dcm
                                             │
                     DVH · radiomics · robustness tables

Each tool runs on its own. RTpipeline runs TotalSegmentator and local nnU-Net models without the portal; the portal serves clinicians and other software without the pipeline. Together, a study can use any model in the portal's catalog for cohort-wide dose and radiomics analysis without installing that model's runtime next to the pipeline, and one GPU host can serve several projects.

The interface between them

The tools exchange data through one documented contract:

  • RTsegmentator writes rtsegmentator_manifest.json into every result ZIP: the source series, each model's outcome and exact ROI names, its licence and limitations, and the size and SHA-256 of every file. The format is a published JSON Schema (rtsegmentator.result-manifest, version 1.0).
  • RTpipeline ships RTsegmentator's standard-library client unchanged (rtpipeline/segmenters/rtsegmentator_client.py) and imports a package only after verifying it and checking that every RTSTRUCT references the course's planning CT.

Configuration, checks and output layout are described in RTsegmentator models.

Running both

deploy/suite/ contains a Docker Compose file that starts the RTsegmentator portal next to the RTpipeline Web UI on one host, with the pipeline configured to reach the portal by its service name. The portal needs an NVIDIA GPU and licensed model assets; see its deployment guide.