--- license: apache-2.0 pipeline_tag: image-segmentation tags: - medical-imaging - image-segmentation - nnunet - nnunetv2 - glioblastoma - brain-tumor - mri --- # SegCFB-GBM A nnU-Netv2 segmentation model for automatic delineation of the Gross Tumor Volume (GTV) in glioblastoma (GBM), from T1-Gadolinium and T2-FLAIR MRI. It was used to complete the longitudinal GTV annotations of the [CFB-GBM v2.0](https://doi.org/10.7937/v9pn-2f72) dataset. ## Model description This is a custom nnU-Netv2 (`3d_fullres`) model trained to segment the contrast-enhancing tumor and necrotic core as a single GTV label. It was first pre-trained on BraTS 2021 and then fine-tuned on the ground-truth GTV contours of the CFB-GBM cohort, which gave the best performance among the configurations we evaluated. The model takes **two input channels**: - channel `0000`: T1-Gadolinium (post-contrast T1) - channel `0001`: T2-FLAIR These two modalities were chosen because they are the most consistently available pair across the CFB-GBM cohort. The exact channel assignment is defined in the bundled `dataset.json`; please verify it there before running inference. ## Intended use The model is intended for research on glioblastoma imaging: completing or bootstrapping GTV annotations, longitudinal tumor tracking, disease progression modelling, and as a segmentation baseline. It is **not** a certified medical device and must not be used for clinical decision-making without expert review. ## Input requirements To match the conditions the model was trained under, inputs should be: - Co-registered so that the T1-Gd and T2-FLAIR volumes share the same space (the CFB-GBM preprocessing used rigid registration to the T1-Gd reference). - Skull-stripped (brain-extracted). - Provided as `.nii.gz`, with the two modalities named `CASE_0000.nii.gz` (T1-Gd) and `CASE_0001.nii.gz` (T2-FLAIR). nnU-Net applies its own intensity normalization internally, so no manual normalization is required. ## Training data - **Pre-training:** BraTS 2021. The BraTS enhancing tumor (ET) and necrotic core (NCR) labels were merged into a single GTV label; the peritumoral edema (ED) label was excluded, as it is not part of a GTV delineation and typically needs additional modalities. - **Fine-tuning:** ground-truth GTV contours from the CFB-GBM cohort (patients treated at Centre François Baclesse under the standard Stupp protocol). ## Performance Test-set overlap against the original manual GTV contours: | Training configuration | Dice (test set) | | --- | --- | | BraTS 2021 only | 0.6128 | | CFB-GBM only | 0.7860 | | **BraTS 2021 + CFB-GBM (this model)** | **0.8031** | ## Repository structure The repo contains a standard nnU-Netv2 results folder with all five folds: ``` Dataset124_CFB-GBM-finetune-train/ └── nnUNetTrainer_100epochs__nnUNetPlans__3d_fullres/ ├── dataset.json ├── plans.json ├── fold_0/checkpoint_best.pth ├── fold_1/checkpoint_best.pth ├── fold_2/checkpoint_best.pth ├── fold_3/checkpoint_best.pth └── fold_4/checkpoint_best.pth ``` ## How to use Install nnU-Netv2 and download this repo: ```bash pip install nnunetv2 hf download AlexLECLERCQ/SegCFB-GBM --local-dir ./SegCFB-GBM ``` Point nnU-Net to the downloaded folder and run inference (ensembling all five folds): ```bash export nnUNet_results="./SegCFB-GBM" nnUNetv2_predict \ -i INPUT_FOLDER \ -o OUTPUT_FOLDER \ -d 124 \ -c 3d_fullres \ -f 0 1 2 3 4 ``` `INPUT_FOLDER` must contain the paired `CASE_0000.nii.gz` (T1-Gd) and `CASE_0001.nii.gz` (T2-FLAIR) files described above. ## Limitations - The model was trained on only two modalities (T1-Gd and T2-FLAIR); accuracy may drop on data with substantially different acquisition or contrast. - The GTV is a single merged label, without the sub-region decomposition (ET, NCR, ED) of the BraTS convention. - Performance was validated on a single-center cohort acquired mostly on Siemens 1.5/3 T scanners; generalization to other centers and scanners has not been characterized. - Outputs are automatic segmentations and require expert verification before any clinical or downstream use. ## License Released under the Apache License 2.0, matching the license of the nnU-Netv2 framework it is built on. ## Citation If you use this model, please cite the associated dataset paper and the CFB-GBM v2.0 dataset. ```bibtex @article{leclercq_cfbgbm_v2, title = {CFB-GBM v2.0: An Augmented Longitudinal Dataset for Multi-Modal Glioblastoma Segmentation, Radiomics, and RANO Progression Tracking}, author = {Leclercq, Alexandre G. and Moreau, Noémie N. and Audebert, Hugo and Nassar, Andros and Cochin, Thomas and Leleu, Thomas and Le Henaff, Loïc and Desmonts, Alexis and Poirier, Yoann and Dubru, Aurélie and Guillemette, Laura and Lecoeur, Pascal and Lemasson, Kévin and Jaudet, Cyril and Bougleux, Sébastien and Hérault, Romain and Brunaud, Carole and Valable, Samuel and Stefan, Dinu and Raboutet, Charlotte and Batalla, Alain and Lacroix, Joëlle and Rouzier, Roman and Corroyer-Dulmont, Aurélien}, journal = {Machine Learning for Biomedical Imaging (MELBA)}, note = {To appear} } @misc{TCIA, title = {Pre and Post Treatment {{MRI}} and Radiotherapy Plans of Patients with Glioblastoma: The {{CFB-GBM}} Cohort ({{CFB-GBM}})}, shorttitle = {Pre and Post Treatment {{MRI}} and Radiotherapy Plans of Patients with Glioblastoma}, author = {Moreau, Noémie N., Leclercq, Alexandre G. and Desmonts, Alexis and Poirier, Yoann and Dubru, Aurélie and Guillemette, Laura and Lecoeur, Pascal and Lemasson, Kévin and Jaudet, Cyril and Brunaud, Carole and Valable, Samuel and Geffrelot, Julien and Stefan, Dinu and Leleu, Thomas and Raboutet, Charlotte and Le Henaff, Loïc and Batalla, Alain and Lacroix, Joëlle and Rouzier, Roman and Corroyer-Dulmont, Aurélien}, year = {2025}, publisher = {The Cancer Imaging Archive}, doi = {10.7937/V9PN-2F72}, url = {https://www.cancerimagingarchive.net/collection/cfb-gbm/}, version = {3} } ``` ## Acknowledgements Data and model developed at Centre François Baclesse (Caen, France). Training scripts are available at . This work was supported by the Région Normandie through the "Booster IA" grant.