Brats 2024 Segmentation Eval

---
name: brats-2024-segmentation-eval
description: Evaluates 3D deep learning models for brain tumor segmentation across three distinct tumor subtypes (pediatric, meningioma, metastasis) using MRI scans. It probes the model's ability to accurately delineate tumor boundaries and generalize across heterogeneous clinical datasets through adaptive post-processing and model ensembling. Use when the user wants to benchmark on BraTS 2024 (PED, MEN-RT, MET), or asks about evaluating this task. Reports lesion-wise Dice score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2412.04094
  bibtex_key: [REDACTED]
  confidence: high
---

# brats-2024-segmentation-eval

> Magnetic Resonance Imaging Feature-Based Subtyping and Model Ensemble for Enhanced Brain Tumor Segmentation — Jiang et al. (2024) (arXiv:2412.04094, 2024)

## What this evaluates

Evaluates 3D deep learning models for brain tumor segmentation across three distinct tumor subtypes (pediatric, meningioma, metastasis) using MRI scans. It probes the model's ability to accurately delineate tumor boundaries and generalize across heterogeneous clinical datasets through adaptive post-processing and model ensembling.

## Datasets

- **BraTS 2024 (PED, MEN-RT, MET)** — total ?; splits: validation (-1), test (-1)

## Metrics

- `lesion-wise Dice score` **(primary)** — range: [0, 1]
  - Computed per lesion instance rather than per voxel: 2 * |A ∩ B| / (|A| + |B|), where A and B are the predicted and ground truth lesion masks after instance matching.
- `95th percentile lesion-wise Hausdorff distance` — range: other
  - Measures the 95th percentile of the maximum surface distance between predicted and ground truth lesion boundaries, computed per lesion instance to reduce sensitivity to extreme outliers.

## Input / output format

**Input**: 3D multi-sequence MRI volumes (e.g., T1, T1ce, T2, FLAIR) with corresponding lesion segmentation masks for training and validation.

**Output**: 3D segmentation masks predicting tumor regions (e.g., whole tumor, enhancing tumor, edema) per instance.

## Scoring recipe

```python
def compute_lesion_metrics(pred_mask, gt_mask):
    pred_lesions = get_connected_components(pred_mask)
    gt_lesions = get_connected_components(gt_mask)
    dice_vals, hd95_vals = [], []
    for gt_comp in gt_lesions:
        best_pred = max(pred_lesions, key=lambda p: iou(p, gt_comp), default=None)
        if best_pred is not None:
            dice_vals.append(2 * np.sum(best_pred & gt_comp) / (np.sum(best_pred) + np.sum(gt_comp) + 1e-6))
            hd95_vals.append(hausdorff_distance(best_pred, gt_comp, percentile=95))
    return np.mean(dice_vals), np.mean(hd95_vals)
```

## Common pitfalls

- Metrics are computed lesion-wise (instance-level) rather than voxel-wise, requiring correct matching of predicted and ground truth lesions before scoring.
- Evaluation is conducted on a centralized platform (Synapse) with no local access to test ground truth, preventing local metric calculation, threshold tuning, or post-hoc analysis.
- Hausdorff distance is reported at the 95th percentile specifically to mitigate sensitivity to boundary noise and extreme outliers, which differs from the standard HD100 metric.

## Evidence (verbatim from paper)

> The models were assessed for each of the regions using the lesion-wise Dice score and the 95th percentile lesion-wise Hausdorff distance.

## Citation

```bibtex
@misc{jiang2024bratsensemble,
  title={Magnetic Resonance Imaging Feature-Based Subtyping and Model Ensemble for Enhanced Brain Tumor Segmentation},
  author={Jiang et al. (2024)},
  year={2024},
  note={arXiv:2412.04094}
}
```

- arXiv: 2412.04094