Brats 2023 Segmentation Eval

---
name: brats-2023-segmentation-eval
description: Evaluates the ability of deep learning models to accurately segment diverse brain tumor sub-regions (enhancing tumor, tumor core, whole tumor) across adult gliomas, pediatric tumors, and sub-Saharan African populations using MRI scans. Use when the user wants to benchmark on BraTS 2023 PED, BraTS 2023 SSA, BraTS-GLI (GLA), or asks about evaluating this task. Reports DSC.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2412.08240
  bibtex_key: [REDACTED]
  confidence: high
---

# brats-2023-segmentation-eval

> Unified HT-CNNs Architecture: Transfer Learning for Segmenting Diverse Brain Tumors in MRI from Gliomas to Pediatric Tumors — Zeineldin et al. (2024) (arXiv:2412.08240, 2024)

## What this evaluates

Evaluates the ability of deep learning models to accurately segment diverse brain tumor sub-regions (enhancing tumor, tumor core, whole tumor) across adult gliomas, pediatric tumors, and sub-Saharan African populations using MRI scans.

## Datasets

- **BraTS 2023 PED** — total ?; splits: validation (-1)
- **BraTS 2023 SSA** — total ?; splits: validation (-1)
- **BraTS-GLI (GLA)** — total ?; splits: validation (-1)

## Metrics

- `DSC` **(primary)** — range: [0, 1]
  - Dice Similarity Coefficient: 2 * |A ∩ B| / (|A| + |B|), where A and B are the prediction and ground truth masks. Values range from 0 to 1, with 1 indicating perfect overlap.
- `HD95` — range: other
  - 95th percentile of the Hausdorff Distance between the boundaries of the prediction and ground truth masks. Computed as the 95th percentile of all pairwise distances between surface points to reduce outlier sensitivity.

## Input / output format

**Input**: Multimodal MRI scans of brain tumors.

**Output**: Pixel-wise segmentation masks for three tumor sub-regions: Enhancing Tumor (ET), Tumor Core (TC), and Whole Tumor (WT).

## Scoring recipe

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

def compute_hd95(pred_mask, gt_mask):
    from scipy.ndimage import distance_transform_edt
    # Compute boundary voxels
    boundary_pred = (pred_mask != 0) & (distance_transform_edt(pred_mask == 0) == 1)
    boundary_gt = (gt_mask != 0) & (distance_transform_edt(gt_mask == 0) == 1)
    # Compute distances between boundary points and take 95th percentile
    dists = distance_transform_edt(boundary_pred == 0)
    hd95_val = np.percentile(dists[boundary_gt], 95)
    return hd95_val
```

## Common pitfalls

- Metrics are computed per sub-region (ET, TC, WT) and then averaged, rather than on a single combined tumor mask.
- HD95 uses the 95th percentile of boundary distances to mitigate extreme outliers, which differs from the standard 100th percentile Hausdorff Distance.

## Evidence (verbatim from paper)

> Standard evaluation metrics, typically DSC and HD95 were used to evaluate segmentation quality. Models pre-trained on Adult Glioma segmentation tasks were fine-tuned on the PED and SSA datasets, showing marked improvements in segmentation performance. The results in Table 3 and Table 4 indicated higher DSC and lower HD95 metrics compared to models trained from scratch, underscoring the benefits of our transfer learning approach.

## Citation

```bibtex
@misc{zeineldin2024unifiedhtcnn,
  title={Unified HT-CNNs Architecture: Transfer Learning for Segmenting Diverse Brain Tumors in MRI from Gliomas to Pediatric Tumors},
  author={Zeineldin et al. (2024)},
  year={2024},
  note={arXiv:2412.08240}
}
```

- arXiv: 2412.08240