Brats2023 Seg Eval

---
name: brats2023-seg-eval
description: Evaluates deep learning models for multi-parametric MRI brain tumor segmentation across three clinical tasks: pediatric tumors, meningiomas, and metastases. It probes the model's ability to accurately delineate tumor sub-regions (enhancing tumor, tumor core, whole tumor) using lesion-wise overlap and boundary distance metrics. Use when the user wants to benchmark on BraTS-PED, BraTS-MEN, BraTS-MET, or asks about evaluating this task. Reports Lesion-wise Dice.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2409.08232
  bibtex_key: capellanmartin2024modelensemble
  confidence: high
---

# brats2023-seg-eval

> Model Ensemble for Brain Tumor Segmentation in Magnetic Resonance Imaging — Capellán-Martín et al. (2024) (arXiv:2409.08232, 2024)

## What this evaluates

Evaluates deep learning models for multi-parametric MRI brain tumor segmentation across three clinical tasks: pediatric tumors, meningiomas, and metastases. It probes the model's ability to accurately delineate tumor sub-regions (enhancing tumor, tumor core, whole tumor) using lesion-wise overlap and boundary distance metrics.

## Datasets

- **BraTS-PED** — total ?; splits: validation (-1), test (-1)
- **BraTS-MEN** — total ?; splits: validation (-1), test (-1)
- **BraTS-MET** — total ?; splits: validation (-1), test (-1)

## Metrics

- `Lesion-wise Dice` **(primary)** — range: [0, 1]
  - Dice = 2 * |Prediction ∩ Ground Truth| / (|Prediction| + |Ground Truth|). Computed per lesion instance across ET, TC, and WT sub-regions.
- `HD95` — range: mm
  - 95th percentile of the Hausdorff distance between the prediction and ground truth surface boundaries, measured in millimeters.

## Input / output format

**Input**: Multi-parametric MRI volumes with corresponding lesion segmentation masks.

**Output**: Binary or multi-class segmentation masks for each tumor sub-region (ET, TC, WT) per lesion.

## Scoring recipe

```python
def compute_metrics(pred_mask, gt_mask):
    intersection = np.sum(pred_mask & gt_mask)
    dice = 2.0 * intersection / (np.sum(pred_mask) + np.sum(gt_mask))
    dists = hausdorff_distance(pred_mask, gt_mask)
    hd95 = np.percentile(dists, 95)
    return dice, hd95
```

## Common pitfalls

- Metrics are computed lesion-wise (per-instance) rather than whole-scan, requiring correct lesion identification and matching.
- The challenge platform automatically computes metrics on the test set without providing ground truth to participants during submission.
- Post-processing with cross-validated thresholds significantly impacts final scores and must be applied consistently across validation and test sets.

## Evidence (verbatim from paper)

> Lesion-wise (LW) Dice coefficients and 95% Hausdorff distances (HD95) were computed for enhancing tumor (ET), tumor core (TC), and whole tumor (WT), respectively.

## Citation

```bibtex
@misc{capellanmartin2024modelensemble,
  title={Model Ensemble for Brain Tumor Segmentation in Magnetic Resonance Imaging},
  author={Capellán-Martín et al. (2024)},
  year={2024},
  note={arXiv:2409.08232}
}
```

- arXiv: 2409.08232