Brats 2013 Eval

---
name: brats-2013-eval
description: Evaluates interactive brain tumor segmentation models by training and testing on a single patient's MRI data to assess within-brain generalization. It measures voxel-wise classification accuracy across different tumor sub-regions using sparse manual labels. Use when the user wants to benchmark on MICCAI-BRATS 2013, or asks about evaluating this task. Reports Dice.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1510.01344
  bibtex_key: havaei2015withinbrain
  confidence: high
---

# brats-2013-eval

> Within-Brain Classification for Brain Tumor Segmentation — Havaei et al. (2015) (arXiv:1510.01344, 2015)

## What this evaluates

Evaluates interactive brain tumor segmentation models by training and testing on a single patient's MRI data to assess within-brain generalization. It measures voxel-wise classification accuracy across different tumor sub-regions using sparse manual labels.

## Datasets

- **MICCAI-BRATS 2013** — total 40; splits: train (30), test (10)

## Metrics

- `Dice` **(primary)** — range: [0, 1]
  - Dice(P,T) = 2|P1∧T1|/(|P1|+|T1|), where P and T are predicted and ground truth voxel sets for a given category. Computed for Complete, Core, and Enhancing regions.
- `Sensitivity` — range: [0, 1]
  - Sensitivity(P,T) = |P1∧T1|/|T1|, measuring the proportion of actual positive voxels correctly identified.
- `Specificity` — range: [0, 1]
  - Specificity(P,T) = |P0∧T0|/|T0|, measuring the proportion of actual negative voxels correctly identified.

## Input / output format

**Input**: 3D MRI volumes (T1C, T2, Flair modalities) with spatial coordinates (i,j,k) per voxel. Sparse manual labels from two 2D slices per class, with voxels outside the skull ignored.

**Output**: 3D voxel-wise segmentation map assigning each voxel to a tumor class (necrosis, edema, enhancing) or healthy tissue.

## Scoring recipe

```python
def compute_metrics(pred_mask, gt_mask):
    intersection = np.sum(pred_mask & gt_mask)
    dice = 2 * intersection / (np.sum(pred_mask) + np.sum(gt_mask))
    sensitivity = intersection / np.sum(gt_mask)
    neg_pred = ~pred_mask
    neg_gt = ~gt_mask
    specificity = np.sum(neg_pred & neg_gt) / np.sum(neg_gt)
    return dice, sensitivity, specificity
```

## Common pitfalls

- Training and testing occur on the same brain (within-brain generalization), not across different patients.
- Only T1C, T2, and Flair modalities are used; T1 is explicitly excluded.
- Hyperparameters are tuned per-brain via cross-validation, not fixed globally.

## Evidence (verbatim from paper)

> All our experiments were conducted on real patient data obtained from the brain tumor segmentation challenge dataset (Farahani et al. [5]) as part of the MICCAI conference. This dataset contains 30 patient subjects (20 high grade and 10 low grade tumors) for training and 10 (all high grade tumors) for testing. The quantitative results for each method was obtained from the BRATS online evaluation system, which provides Dice, Specificity and Sensitivity as measures of performance. These measures are defined as follows: Dice(P,T) = 2|P1∧T1|/(|P1|+|T1|), Sensitivity(P,T) = |P1∧T1|/|T1|, Specificity(P,T) = |P0∧T0|/|T0|

## Citation

```bibtex
@misc{havaei2015withinbrain,
  title={Within-Brain Classification for Brain Tumor Segmentation},
  author={Havaei et al. (2015)},
  year={2015},
  note={arXiv:1510.01344}
}
```

- arXiv: 1510.01344