Brats2017 Segmentation Eval

---
name: brats2017-segmentation-eval
description: Evaluates the ability of 3D U-Net architectures to segment brain tumors (enhancing tumor, whole tumor, and tumor core) from multimodal MRI scans. It specifically probes how well lesion prior information (VOI maps) can be fused with imaging data to improve volumetric segmentation accuracy and boundary precision. Use when the user wants to benchmark on BraTS 2017, or asks about evaluating this task. Reports DSC (Dice Similarity Coefficient).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1907.00281
  bibtex_key: [REDACTED]
  confidence: high
---

# brats2017-segmentation-eval

> Improving 3D U-Net for Brain Tumor Segmentation by Utilizing Lesion Prior — Kao et al. (2019) (arXiv:1907.00281, 2019)

## What this evaluates

Evaluates the ability of 3D U-Net architectures to segment brain tumors (enhancing tumor, whole tumor, and tumor core) from multimodal MRI scans. It specifically probes how well lesion prior information (VOI maps) can be fused with imaging data to improve volumetric segmentation accuracy and boundary precision.

## Datasets

- **BraTS 2017** — total ?; splits: train (285), val (-1)

## Metrics

- `DSC (Dice Similarity Coefficient)` **(primary)** — range: [0, 1]
  - 2 * |A ∩ B| / (|A| + |B|), where A and B are the predicted and ground truth binary masks for a tumor sub-region.
- `H95 (95th percentile Hausdorff Distance)` — range: mm
  - 95th percentile of the maximum distances from any point in the predicted boundary to the nearest point in the ground truth boundary.

## Input / output format

**Input**: Multimodal MRI volumes (FLAIR, T1, T1ce, T2) concatenated with a probabilistic Volume-of-Interest (VOI) map indicating lesion likelihood.

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

## Scoring recipe

```python
def compute_metrics(pred, gt):
    intersection = np.logical_and(pred, gt).sum()
    union = pred.sum() + gt.sum()
    dsc = 2 * intersection / union if union > 0 else 0.0
    h95 = percentile(hausdorff_distance(pred_boundary, gt_boundary), 95)
    return dsc, h95
# Evaluated on BraTS 2017 validation set; results reported as mean across subjects.
```

## Common pitfalls

- DSC and H95 have opposite optimization directions (higher DSC is better, lower H95 is better), which can be confusing when comparing results.
- The BraTS challenge uses an official online evaluation server; local implementations of H95 or DSC may yield slightly different values due to preprocessing or boundary definition differences.
- Tumor Core (TC) is defined as the union of necrosis/non-enhancing tumor and enhancing tumor, not just the enhancing tumor.

## Evidence (verbatim from paper)

> BraTS 2017 validation set is used to evaluate the performance of these networks. The quantitative results are shown in Table 1. Higher DSC and lower H95 indicate better segmentation performance. These results are given by the official online evaluation website. Results are reported as mean.

## Citation

```bibtex
@misc{kao2019improving,
  title={Improving 3D U-Net for Brain Tumor Segmentation by Utilizing Lesion Prior},
  author={Kao et al. (2019)},
  year={2019},
  note={arXiv:1907.00281}
}
```

- arXiv: 1907.00281