# Brats T1t2 Seg Eval

> Evaluates a model's ability to perform unsupervised domain adaptation for brain tumor segmentation, specifically transferring segmentation capabilities from T1-weighted MRI scans to T2-weighted MRI scans without target labels. Use when the user wants to benchmark on BraTS'19, or asks about evaluating this task. Reports DSC.

- Skill: `qhjqhj00/brats-t1t2-seg-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/brats-t1t2-seg-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/brats-t1t2-seg-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/brats-t1t2-seg-eval

---


# brats-t1t2-seg-eval

> Cross-Modality Brain Tumor Segmentation via Bidirectional Global-to-Local Unsupervised Domain Adaptation — He et al. (2021) (arXiv:2105.07715, 2021)

## What this evaluates

Evaluates a model's ability to perform unsupervised domain adaptation for brain tumor segmentation, specifically transferring segmentation capabilities from T1-weighted MRI scans to T2-weighted MRI scans without target labels.

## Datasets

- **BraTS'19** — total ?; splits: test (-1)

## Metrics

- `DSC` **(primary)** — range: percent
  - Dice Similarity Coefficient. Computed per class (WT, TC, ET) as 2|P∩G|/(|P|+|G|), then averaged across classes. Reported as a percentage.
- `HD95` — range: mm
  - 95th percentile of the Hausdorff Distance between predicted and ground-truth segmentation boundaries. Measured in millimeters to reduce sensitivity to outliers.

## Input / output format

**Input**: 3D MRI volume (T1-weighted for source domain, T2-weighted for target domain).

**Output**: 3D segmentation mask with three classes: Whole Tumor (WT), Tumor Core (TC), Enhancing Tumor (ET).

## Scoring recipe

```python
def compute_dsc(pred, gt):
    intersection = np.sum(pred * gt)
    return 2.0 * intersection / (np.sum(pred) + np.sum(gt))

def compute_hd95(pred, gt, spacing=1.0):
    from scipy.ndimage import distance_transform_edt
    dist_pred = distance_transform_edt(1 - pred) * spacing
    dist_gt = distance_transform_edt(1 - gt) * spacing
    boundary_dists = np.concatenate([dist_pred[pred==1], dist_gt[gt==1]])
    return np.percentile(boundary_dists, 95)
```

## Common pitfalls

- Metrics are computed separately for three tumor sub-regions (WT, TC, ET) before averaging.
- HD95 uses the 95th percentile to mitigate extreme boundary errors common in medical segmentation.
- Evaluation assumes no target-domain labels are used during training (unsupervised domain adaptation).

## Evidence (verbatim from paper)

> TABLE I: Comparison our model with the state-of-the-art methods in DSC (%) for the task of T1 to T2. The best results are highlighted in bold. TABLE II: Performance comparison with the state-of-the-art methods in HD95 (mm) for the task of T1 to T2.

## Citation

```bibtex
@misc{he2021crossmodality,
  title={Cross-Modality Brain Tumor Segmentation via Bidirectional Global-to-Local Unsupervised Domain Adaptation},
  author={He et al. (2021)},
  year={2021},
  note={arXiv:2105.07715}
}
```

- arXiv: 2105.07715