Skip to content

Latest commit

 

History

History
82 lines (63 loc) · 3.68 KB

File metadata and controls

82 lines (63 loc) · 3.68 KB

Model Card — SwinUNETR v2 Brain Tumour Segmentation

Model Details

Field Value
Architecture SwinUNETR v2 (Swin Transformer + U-Net decoder)
Framework MONAI 1.x + PyTorch
Task 3-class brain tumour segmentation (TC, WT, ET)
Input 4-channel multi-modal MRI: FLAIR, T1, T1ce, T2 at 1 mm³ isotropic
Output 3 independent binary masks (sigmoid, not softmax)
Training precision bfloat16 AMP
Checkpoint W&B artifact swin-unetr-brats2021/best_checkpoint_fold0:v10

Training Data

Field Value
Dataset BraTS 2021 Task 1 (Multimodal Brain Tumour Segmentation Challenge)
Source Kaggle — dschettler8845/brats-2021-task1
Total cases 1 251 subjects
Split 875 train / 188 val / 188 test
Label convention Label 1 = Tumour Core (TC); Label 2 = Whole Tumour (WT); Label 3 = Enhancing Tumour (ET)
Anonymisation All BraTS cases are de-identified per challenge protocol
Demographics Not available — BraTS does not release patient metadata

Evaluation Results (held-out test set, 188 cases)

Metric Value
Mean Dice 0.8961
Tumour Core Dice 0.8787
Whole Tumour Dice 0.9282
Enhancing Tumour Dice 0.8814
Mean HD95 (mm) 5.62
TC HD95 (mm) 6.01
WT HD95 (mm) 5.85
ET HD95 (mm) 4.98

The MONAI SwinUNETR v2 paper reports ~0.875 mean Dice on BraTS 2021 validation. This model exceeds that baseline.

Intended Use

Intended: Research, education, and demonstration of deep-learning-based medical image segmentation.

Not intended: Clinical diagnosis, treatment planning, or any patient-facing medical decision support. This model has not been validated on out-of-distribution scanners, acquisition protocols, or patient populations beyond BraTS 2021.

Limitations and Biases

  • Trained exclusively on BraTS 2021 data (adult glioma). Performance on paediatric tumours, metastases, or other pathologies is unknown.
  • BraTS ground-truth labels were created by expert radiologists but contain inter-annotator variability, particularly at tumour boundaries.
  • Scanner heterogeneity in BraTS may not represent the full distribution of clinical MRI scanners.
  • The model does not produce calibrated uncertainty estimates — confidence scores are mean sigmoid activations, not probabilities.
  • No evaluation has been performed on populations underrepresented in BraTS (e.g. low-resource healthcare settings).

Ethical Considerations

  • MRI data uploaded to the demo is processed in-memory and not stored or logged.
  • The system displays a mandatory disclosure notice (EU AI Act transparency requirement) before any user interaction.
  • All outputs carry a disclaimer: "AI research output only — NOT a clinical diagnosis."

EU AI Act Classification

Under the EU AI Act (Regulation 2024/1689), AI systems used as medical devices or for purposes related to medical diagnosis are listed in Annex III as high-risk AI systems. This model is deployed for research and demonstration purposes only. It has not undergone conformity assessment under Article 43 and must not be placed on the market or put into service as a medical device.

Citation

@inproceedings{tang2022self,
  title={Self-supervised pre-training of swin transformers for 3d medical image analysis},
  author={Tang, Yucheng and others},
  booktitle={CVPR},
  year={2022}
}

@article{baid2021rsna,
  title={The RSNA-ASNR-MICCAI BraTS 2021 Benchmark},
  author={Baid, Ujjwal and others},
  journal={arXiv preprint arXiv:2107.02314},
  year={2021}
}