Brats2021 Segmentation Eval

---
name: brats2021-segmentation-eval
description: Evaluates 3D brain tumor segmentation accuracy and robustness under missing MRI modalities using multi-modal MRI scans. It probes a model's ability to delineate tumor sub-regions while maintaining calibration and stability when contrast sequences are corrupted or absent. Use when the user wants to benchmark on BraTS 2021, or asks about evaluating this task. Reports Dice Score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2601.17504
  bibtex_key: zhou2026bmdsnet
  confidence: high
---

# brats2021-segmentation-eval

> BMDS-Net: A Bayesian Multi-Modal Deep Supervision Network for Robust Brain Tumor Segmentation — Zhou et al. (2026) (arXiv:2601.17504, 2026)

## What this evaluates

Evaluates 3D brain tumor segmentation accuracy and robustness under missing MRI modalities using multi-modal MRI scans. It probes a model's ability to delineate tumor sub-regions while maintaining calibration and stability when contrast sequences are corrupted or absent.

## Datasets

- **BraTS 2021** — total 1251; splits: train (-1), val (-1), test (-1)

## Metrics

- `Dice Score` **(primary)** — range: [0, 1]
  - Dice similarity coefficient computed per tumor sub-region (Whole Tumor, Tumor Core, Enhancing Tumor) between predicted and ground truth segmentation masks.
- `HD95` — range: other
  - 95th percentile of the Hausdorff distance between predicted and ground truth boundaries, measured in millimeters.
- `ECE` — range: [0, 1]
  - Expected Calibration Error measuring the alignment between predicted confidence probabilities and actual accuracy.

## Input / output format

**Input**: Multi-modal 3D MRI volumes (T1, T1ce, T2, FLAIR) preprocessed with z-score normalization and randomly cropped to 128x128x128 voxels during training.

**Output**: Voxel-wise segmentation masks for three tumor sub-regions (WT, TC, ET) and voxel-wise uncertainty estimates.

## Scoring recipe

```python
def compute_dice(pred, gt):
    intersection = np.sum(pred * gt)
    union = np.sum(pred) + np.sum(gt)
    return 2.0 * intersection / (union + 1e-5)

def compute_hd95(pred, gt, spacing=1.0):
    dist = hausdorff_distance(pred, gt)
    return np.percentile(dist, 95) * spacing

def compute_ece(confidence, accuracy, n_bins=15):
    bin_boundaries = np.linspace(0, 1, n_bins + 1)
    ece = 0.0
    for i in range(n_bins):
        mask = (confidence >= bin_boundaries[i]) & (confidence < bin_boundaries[i+1])
        if np.sum(mask) > 0:
            bin_acc = np.mean(accuracy[mask])
            bin_conf = np.mean(confidence[mask])
            ece += np.sum(mask) * abs(bin_acc - bin_conf)
    return ece / len(accuracy)
```

## Common pitfalls

- The test set is extremely small (5% of 1,251 ≈ 62 scans), which can cause high variance in reported metrics and limit statistical power.
- Missing-modality robustness is evaluated on the validation set rather than the official BraTS test set, making direct comparison with public benchmarks difficult.
- HD95 is reported in millimeters, but the exact voxel spacing/resolution used for conversion is not explicitly stated, which affects cross-study comparability.

## Evidence (verbatim from paper)

> BMDS-Net achieves competitive Dice scores comparable to the highly optimized nnU-Net. Crucially, compared to the vanilla Swin UNETR, our method demonstrates consistent improvements across all tumor sub-regions, validating the efficacy of our proposed modules.

## Citation

```bibtex
@misc{zhou2026bmdsnet,
  title={BMDS-Net: A Bayesian Multi-Modal Deep Supervision Network for Robust Brain Tumor Segmentation},
  author={Zhou et al. (2026)},
  year={2026},
  note={arXiv:2601.17504}
}
```

- arXiv: 2601.17504