Glioblastoma Segmentation (CFB-GBM Cohort)
This model is a 3D Full Resolution nnU-Net trained to segment the Gross Tumor Volume (GTV) of Glioblastoma patients.
It has been trained on the CFB-GBM Cohort (Centre François Baclesse - Glioblastoma), a highly curated dataset hosted on The Cancer Imaging Archive (TCIA), representing real-world clinical heterogeneity.
⚠️ Critical Input Requirements
To achieve correct segmentation, inputs MUST match the preprocessing pipeline of the CFB-GBM dataset used for training.
| Channel | Modality | Preprocessing Requirement |
|---|---|---|
| 0000 | T1-Weighted Gado (T1Gd) | Skull-stripped & Reference space |
| 0001 | T2-FLAIR | Skull-stripped & Co-registered to T1Gd |
Note: The model expects brain-extracted images. Using raw DICOM/NIfTI with the skull intact will likely result in poor performance or hallucinations.
Dataset & Training Context
This model was trained using the CFB-GBM dataset (v1.0), comprising 264 glioblastoma patients from clinical routine (2017-2023).
- Data Source: The Cancer Imaging Archive (CFB-GBM Collection)
- Population: Real-world clinical routine (not a clinical trial cohort), ensuring robustness to heterogeneous acquisitions.
- Ground Truth: GTV manually delineated by expert radiation oncologists using treatment planning software (Raystation/Eclipse/Precision).
Model Performance
- Architecture: 3D U-Net (Full Resolution)
- Validation Dice (Fold 0): ~0.85
- Target Spacing:
[1.0, 0.5, 0.5]mm (Optimized for high-anisotropy clinical scans) - Training Hardware: NVIDIA RTX 5090 (with specific compile flags)
Visual Results
3D Volumetric Reconstruction
Usage Instructions (RTX 5090 / Modern GPUs)
Engineering Note: Standard nnU-Net inference may crash on newer GPUs (RTX 4090/5090) due to torch compile/Triton backend incompatibilities.
1. Environment Setup
conda create env -n nnUnetSeg python=3.11
conda activate nnUnetSeg
pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130
pip install nnunetv2
2. Inference Command
Ensure your files are named strictly (e.g., Case_0000.nii.gz for T1Gd and Case_0001.nii.gz for FLAIR).
# Setup paths
export nnUNet_raw="./path/to/raw"
export nnUNet_results="./path/to/results"
# ENGINEERING HACKS FOR STABILITY (RTX 5090)
export nnUNet_compile=f
export TRITON_DISABLE=1
# Run Inference (Fold 0)
nnUNetv2_predict \
-i ${nnUNet_raw}/images_inference \
-o ${nnUNet_raw}/output_inference \
-d 1 \
-c 3d_fullres \
-f 0 \
-chk checkpoint_best.pth
Citation & Acknowledgements
This model is a derivative work based on the CFB-GBM cohort. If you use this model, you must cite the original dataset:
MOREAU, N. N., LECLERCQ, A. G., DESMONTS, A., & CORROYER-DULMONT, A. (2025). > Pre and post treatment MRI and radiotherapy plans of patients with glioblastoma: the CFB-GBM cohort (CFB-GBM) (Version 1) [Data set]. The Cancer Imaging Archive. DOI: 10.7937/V9PN-2F72
Predicted Dataset available here: HuggingFace Dataset