This repository provides the full code and configurations used for our submission to the PANTHER Challenge 2025, focused on automatic pancreatic tumor segmentation on MR-Linac T2-weighted MRI.
Our approach leverages transfer learning, fine-tuning, and ensemble methods to tackle the few-shot setting (50 annotated cases).
| Rank | Team / User | Method | DSC (↑) | NSD (↑) | HD95 [mm] (↓) | ASD [mm] (↓) | Vol. [mm³] (↑) |
|---|---|---|---|---|---|---|---|
| 1st | MIC-DKFZ | Bauchspeichel-Doppel-Düse | 0.5289 | 0.6999 | 23.0110 | 5.1319 | 17163.5753 (4) |
| 2nd | BreizhSeg | docker_Task2 | 0.4910 | 0.6736 | 25.2634 (4) | 6.1722 (2) | 16714.9043 (2) |
| 2nd | LiboZhang | Task2Baseline | 0.4814 | 0.6507 | 24.0864 (2) | 7.6881 (3) | 16647.2197 (1) |
| 4th | amparobt9 | PANTHER baseline (MRSegmentator) | 0.4784 | 0.6457 | 24.1735 (3) | 8.1224 (4) | 16846.2910 (3) |
Our approach explored three main strategies:
-
MRI → CT Transfer
- Used SynthRAD2025 models to generate synthetic CT (sCT) from T2 MRI https://github.com/vboussot/Synthrad2025_Task_1.
- Applied pretrained CT tumor segmentation models from Curvas2025 challenge https://github.com/vboussot/CurvasPDACVI.
- Limitation: tumors often blurred or absent in sCT.
-
Direct MRI Fine-tuning
- Fine-tuned a pretrained PANORAMA ResUNetL (winner of CT segmentation) on Panther T2W MRIs.
- Showed competitive performance but high inter-patient variability.
-
SAM 2.1 Fine-tuning with Prompts
- Used method (2) segmentations as prompts for SAM 2.1.
- Fine-tuned SAM on Panther dataset.
- Improvement marginal; SAM struggled with medical tumor characteristics.
We combined the panther curvas with multiple fine-tuned models with STAPLE ensembling, which provided the most robust approach across cases.
- 🔗 Our fine-tuned Panther weights: Hugging Face – VBoussot/Panther
- 🔗 Official baseline model (MRSegmentator v1.2.0): GitHub Release