MedOtter/PSMA-PET-CT-Lesions

PSMA-PET-CT-Lesions

597 whole-body PSMA-PET/CT studies from 378 male patients with suspected or diagnosed prostate carcinoma, acquired at LMU University Hospital, Munich (2014-2022), with manually segmented PSMA-avid tumor lesions. Together with the FDG sibling collection this formed the training data for the autoPET III / IV challenges.

This is the first PET dataset in the MedOtter suite.

Layout

imagesTr/<case_id>_0000.nii.gz   CT   (resampled onto the PET grid)
imagesTr/<case_id>_0001.nii.gz   PET  (SUV, body-mass normalized)
labelsTr/<case_id>.nii.gz        binary lesion mask, uint8, {0, 1}
dataset.json, dataset_fingerprint.json, splits_final.json, psma_metadata.csv   (upstream, verbatim)
cases.jsonl                      added by this mirror -- see below

case_id is psma_<16-hex patient hash>_<YYYY-MM-DD>.

cases.jsonl (added by this mirror)

One row per study joining the upstream metadata with the official fold assignment and a locally measured lesion load:

case_id, patient_id, study_date, ct, pet, mask, has_lesion, n_lesion_voxels, fold, age, scanner, pet_radionuclide, ct_contrast_agent, shape, spacing.

Splits

There is no official train/test split. The upstream splits_final.json is a 5-fold cross-validation split, and it is already patient-grouped -- verified here: zero patient overlap between train and val in all five folds, union exactly 597. The fold column of cases.jsonl records which fold each study is the validation member of, so any fold can be reconstructed exactly.

Grouping warning: a study is NOT a patient. 164 of the 378 patients contribute 2-7 studies. Group on patient_id for any split you construct yourself.

Channel roles -- upstream bug

dataset.json declares "channel_names": {"0": "CT", "1": "CT"}. Channel 1 is PET in SUV, not a second CT. The file is mirrored verbatim; use the ct / pet keys in cases.jsonl instead. (dataset.json's "name" field also still reads PSMA-PET-CT-Lesions_v2 inside the v3 archive.)

Annotation

All PSMA-avid lesions were segmented on the PET images by a single reader with 3 years of hybrid-imaging experience, using circular VOIs with an uptake-threshold pre-segmentation followed by slice-by-slice manual correction (mint Medical, Heidelberg), then validated by two board-certified experts (4 and >10 years). Single annotation tier -- no multi-rater set, no STAPLE, no pseudo-labels.

Provenance

Mirrored from the authors' own NIfTI release: FDAT (University of Tuebingen) record `g27kx-86t35`, PSMA-PET-CT_Lesions_v3.zip (20,588,970,455 bytes). Counts verified independently against the paper, the TCIA NBIA API and psma_metadata.csv: 597 studies / 378 patients / 539 positive / 58 negative, and CT/PET/mask shape+affine equality checked on all 597 studies.

Route matters. The same collection is also on TCIA as DICOM under CC BY 4.0 (10.7937/r7ep-3x37). This mirror follows the FDAT NIfTI release, which is CC BY-NC 4.0 -- so this copy is for non-commercial research use. The TCIA DICOM copy additionally differs: it is de-faced (extended head-region anonymization) and its study dates are shifted to 1997-2005, so TCIA and FDAT cannot be joined by date -- only by the patient hash.

Overlap / leakage notes

• —No patient overlap with the FDG FDG-PET-CT-Lesions (autoPET II, Tuebingen) collection -- different institution, tracer and ID namespace.
• —DAP_Atlas annotates the FDG autoPET CT, not this collection.
• —autoPET III training data = this collection + the FDG collection (597 + 1,014 studies). The challenge itself used the pre-defacing v1 bytes.
• —The autoPET III test set includes 50 PSMA_LMU studies from this same institution, and the paper does not assert they are disjoint from these 597. Treat autoPET III leaderboard comparisons as carrying unquantified overlap risk.
• —FDAT v3 changed the SUV computation (QIBA-aligned); v2 and v3 PET voxel values are not identical. The v2 release carried a mask correction to 5 studies from 3 patients.
• —Not to be confused with DEEP-PSMA (Peter MacCallum, MICCAI 2025) -- a different, disjoint cohort.

License

CC BY-NC 4.0 (per the FDAT record). Non-commercial research use; please attribute the authors and cite the paper below.

Citation

@article{jeblick2026psma,
  title={A whole-body PSMA-PET/CT dataset with manually annotated tumor lesions},
  author={Jeblick, Katharina and Schachtner, Balthasar and Mittermeier, Andreas and
          Dexl, Jakob and Wesp, Philipp and K{\"u}stner, Thomas and Gatidis, Sergios and
          Fr{\"u}h, Marcel and Fabritius, Matthias P. and Herr, Fabian and
          Unterrainer, Lena and Klimek, Konrad and Sheikh, Gabriel and B{\"o}ning, Guido and
          Brendel, Matthias and Ricke, Jens and Werner, Rudolf A. and Gu, Sen and
          Shiyam Sundar, Lalith Kumar and Ingrisch, Michael and Geyer, Thomas and Cyran, Clemens},
  journal={Scientific Data},
  volume={13}, pages={1023}, year={2026},
  doi={10.1038/s41597-026-07821-z}
}