Grand Challenge algorithm for team hazy_PET in the autoPET V (2026) challenge: interactive tumor-lesion segmentation in whole-body FDG / PSMA PET/CT, refined from sparse scribble corrections during inference.
Current stage: baseline (official nnU-Net). The method is under development; this repository will be updated as the algorithm evolves.
The inference container (algorithm/) follows the Grand Challenge interface:
- Input:
/input/images/ct/*.mha,/input/images/pet/*.mha,/input/lesion-clicks.json - Output:
/output/images/tumor-lesion-segmentation/*.mha - Scribbles from
lesion-clicks.jsonare encoded as foreground / background guidance channels; the model (nnU-Net Dataset998, 3D full-resolution) predicts a binary tumor-lesion segmentation, refined over up to 5 interaction steps.
cd algorithm
bash build.sh # builds the Docker image (checks/fetches weights first)
bash test.sh # runs the container on the bundled test interface
bash export.sh # saves a submittable image tarballTrained weights are not included in this repository (large binaries). See
algorithm/check_weights.sh for retrieval; weights for the released version are
distributed via a GitHub Release / Zenodo (link to be added).
Released under the Apache License 2.0 (LICENSE). The container is based on the
official autoPET V baseline by lab-midas — see NOTICE for attribution.
A method paper (preprint) accompanying this submission will be linked here.