| --- |
| license: cc-by-nc-4.0 |
| task_categories: |
| - image-segmentation |
| tags: |
| - medical-imaging |
| - ct |
| - pancreatic-cancer |
| - longitudinal |
| - lesion-tracking |
| - segmentation |
| pretty_name: PanTrack |
| --- |
| |
| # 🥞 *PanTrack*: Longitudinal Pancreatic Cancer Tracking Benchmark |
|
|
| <div align="center"> |
|
|
| [](https://huggingface.co/datasets/mrokuss/PanTrack) |
| [](https://arxiv.org/abs/2605.23118)  |
| [](https://github.com/MIC-DKFZ/LongiSeg)  |
|
|
| </div> |
|
|
| <img src="https://huggingface.co/datasets/mrokuss/PanTrack/resolve/main/PanTrack.jpg" alt="PanTrack Logo"/> |
|
|
| PanTrack is a longitudinal CT benchmark for tumor lesion tracking and segmentation, introduced alongside the paper: |
|
|
| > **Exploiting Longitudinal Context in Clinician-Verified Interactive Lesion Tracking** |
| > Yannick Kirchhoff\*, Maximilian Rokuss\*, et al. — *MICCAI 2026* |
| > [\[Paper\]](https://arxiv.org/abs/2605.23118) [\[Code & Models\]](https://github.com/MIC-DKFZ/LongiSeg) |
|
|
| This dataset was used **exclusively as a held-out out-of-distribution (OOD) test set**. It was never used for training or model selection in any experiments reported in the paper. It serves as a rigorous cross-domain generalization benchmark for lesion tracking models trained on other data (e.g., autoPET/CT IV). |
|
|
| --- |
|
|
| ## Dataset Overview |
|
|
| | Property | Value | |
| |---|---| |
| | Patients | 45 with pancreatic adenocarcinoma | |
| | CT examinations | 161 (2–11 per patient, mean 3.6) | |
| | Acquisition | Portal-venous phase, identical Siemens protocols, single institution | |
| | Lesion types | Pancreatic tumors + hepatic metastases (partially annotated) | |
| | Total annotated lesion instances | 292 (165 pancreas, 124 liver, 3 lymph nodes) | |
| | Tracking pairs (consecutive BL→FU) | 116 | |
| | Vanishing lesion pairs (no FU annotation) | 36 (stored as `null` / `NaN` in tracking.json) | |
|
|
| **Important caveats:** |
| - **Not all liver lesions are annotated.** Only hepatic metastases deemed relevant by the radiologist are included. The absence of a liver annotation does not imply a lesion-free liver. |
| - Lesion **indices are non-consecutive integers** (e.g., 51, 81, 188, 193, 200, …). They are instance label values in the NIfTI masks and serve as stable cross-timepoint identifiers. |
| - Pancreatic lesions characteristically have **fuzzy boundaries and subtle soft-tissue contrast**, making them harder to segment than e.g. melanoma metastases. |
| - The diverse cohort includes patients on long-term stable chemotherapy and those with rapid progression. |
|
|
| --- |
|
|
| ## File Structure |
|
|
| ``` |
| images/ # 161 CT scans, NIfTI format |
| {PatientID}_{YYYYMMDD}_0000.nii.gz |
| |
| labels/ # 161 instance segmentation masks, NIfTI format |
| {PatientID}_{YYYYMMDD}.nii.gz |
| |
| totalseg/ # 161 TotalSegmentator whole-body organ masks |
| {PatientID}_{YYYYMMDD}.nii.gz |
| |
| patients.json # patient → list of scan IDs (chronological order) |
| tracking.json # BL→FU tracking pairs with lesion center points |
| organ_annotations.json # per-scan lesion metadata (organ, anatomical location) |
| recommended_split.json # suggested train/val split for future experimentation |
| ``` |
|
|
| --- |
|
|
| ## Label Format |
|
|
| Labels are **instance segmentation masks** stored as `uint16` NIfTI files. Each non-zero voxel value identifies a specific lesion instance that is **consistent across all timepoints of the same patient**. |
|
|
| - `0` = background |
| - Non-zero integer = lesion instance ID (non-consecutive; e.g. 51, 81, 188–200) |
|
|
| The same lesion in scan `PanTrack_004_20221005` and `PanTrack_004_20230103` will carry the same integer label value. |
|
|
| --- |
|
|
| ## `tracking.json` |
|
|
| Top-level structure: `{ PatientID → [ pair_dict, pair_dict, … ] }`. |
| Each `pair_dict` covers one consecutive baseline→follow-up step and maps `lesion_id → entry`: |
|
|
| ```json |
| "PanTrack_001": [ |
| { |
| "51": { |
| "img_bl": "PanTrack_001_20220111", |
| "img_fu": "PanTrack_001_20220711", |
| "bl_point": [235.77, 186.07, 664.71], // lesion centroid in BL, voxel (x,y,z) |
| "fu_point": [278.76, 165.05, 629.30], // lesion centroid in FU, voxel (x,y,z) |
| // null / NaN = vanishing lesion |
| "fu_point_prop": [230.65, 175.46, 652.75], // uniGradICON-propagated BL→FU point |
| "merged_lesions": [51] // FU label IDs the lesion merged into |
| } |
| } |
| ] |
| ``` |
|
|
| All coordinates are **voxel-space (x, y, z)**, i.e. to convert to array index use `(z, y, x)`. |
|
|
| --- |
|
|
| ## `organ_annotations.json` |
| |
| Maps each scan to a dict of `{ lesion_id → { organ, location, display_text } }`: |
| |
| ```json |
| "PanTrack_004_20230920": { |
| "195": { "organ": "liver", "location": "Leber Segment V", "display_text": "PAM03 Leber Segment V" }, |
| "200": { "organ": "liver", "location": "Leber Segment VI", "display_text": "PAM08 Leber Segment VI" }, |
| "188": { "organ": "pancreas", "location": "Pankreas Kopf", "display_text": "PAT01 Pankreas Kopf" } |
| } |
| ``` |
| |
| Note: the lesion IDs in this file use a different numbering scheme than the label mask instance values. Use `tracking.json` as the primary source for cross-timepoint correspondence. |
| |
| --- |
| |
| ## `recommended_split.json` |
| |
| A single-fold split (109 train / 52 val scans) provided for researchers who wish to fine-tune or train on PanTrack data. **Not used in the original paper** since the full dataset was treated as a zero-shot OOD test set. |
| |
| --- |
| |
| ## 🔗 Also Interesting |
| |
| Related work on longitudinal medical image analysis and lesion tracking: |
| |
| **LesionLocator: Zero-Shot Universal Tumor Segmentation and Tracking in 3D Whole-Body Imaging** *(CVPR 2025)* |
| A framework for zero-shot lesion segmentation and longitudinal tumor tracking in 3D full-body imaging, combining a large-scale lesion dataset, promptable segmentation, and deep-learning-based image registration. |
| [](https://openaccess.thecvf.com/content/CVPR2025/html/Rokuss_LesionLocator_Zero-Shot_Universal_Tumor_Segmentation_and_Tracking_in_3D_Whole-Body_CVPR_2025_paper.html) [\[GitHub\]](https://github.com/MIC-DKFZ/LesionLocator) |
| |
| **Longitudinal Segmentation of MS Lesions via Temporal Difference Weighting** *(MICCAI 2024)* |
| Introduces a Difference Weighting Block that explicitly incorporates temporal differences between baseline and follow-up MRI scans for superior MS lesion segmentation and detection. |
| [](https://arxiv.org/abs/2409.13416) [\[GitHub\]](https://github.com/MIC-DKFZ/Longitudinal-Difference-Weighting) |
| |
| --- |
| |
| ## Citation |
| |
| If you use *PanTrack*, please cite: |
| |
| ```bibtex |
| @inproceedings{kirchhoff2026pantrack, |
| title = {Exploiting Longitudinal Context in Clinician-Verified Interactive Lesion Tracking}, |
| author = {Kirchhoff, Yannick and Rokuss, Maximilian and Mertens, Daniel Philipp and |
| F{\"u}ller, David and Hamm, Benjamin and Schreyer, Andreas and |
| Ritter, Oliver and Maier-Hein, Klaus}, |
| booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026}, |
| year = {2026} |
| } |
| @inproceedings{rokuss2024longitudinal, |
| title={Longitudinal segmentation of MS lesions via temporal Difference Weighting}, |
| author={Rokuss, Maximilian R and Kirchhoff, Yannick and Roy, Saikat and Kovacs, Balint and Ulrich, Constantin and Wald, Tassilo and Zenk, Maximilian and Denner, Stefan and Isensee, Fabian and Vollmuth, Philipp and Kleesiek, Jens and Maier-Hein, Klaus}, |
| booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention}, |
| pages={64--74}, |
| year={2024}, |
| organization={Springer} |
| } |
| ``` |
| |
