Brats2015 Segmentation Eval

---
name: brats2015-segmentation-eval
description: Evaluates a model's ability to automatically segment brain tumor subtypes (complete, core, enhancing) from multimodal MRI scans. It specifically probes performance on high-grade gliomas (HGG) versus low-grade gliomas (LGG), highlighting challenges with ambiguous boundaries and lack of contrast enhancement. Use when the user wants to benchmark on BRATS 2015, or asks about evaluating this task. Reports Dice Similarity Coefficient (DSC).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1705.03820
  bibtex_key: dong2017automatic
  confidence: high
---

# brats2015-segmentation-eval

> Automatic Brain Tumor Detection and Segmentation Using U-Net Based Fully Convolutional Networks — Dong et al. (2017) (arXiv:1705.03820, 2017)

## What this evaluates

Evaluates a model's ability to automatically segment brain tumor subtypes (complete, core, enhancing) from multimodal MRI scans. It specifically probes performance on high-grade gliomas (HGG) versus low-grade gliomas (LGG), highlighting challenges with ambiguous boundaries and lack of contrast enhancement.

## Datasets

- **BRATS 2015** — total ?; splits: train (-1), test (-1)

## Metrics

- `Dice Similarity Coefficient (DSC)` **(primary)** — range: [0, 1]
  - Computes the overlap between predicted segmentation mask and ground truth mask: DSC = 2 * |intersection| / (|prediction| + |ground_truth|). Values range from 0 to 1, where 1 indicates perfect overlap.

## Input / output format

**Input**: Stacked multimodal MRI channels (T1, T2, FLAIR, and T1c) for each patient case.

**Output**: Binary segmentation masks for three tumor regions: complete tumor, core tumor, and enhancing tumor.

## Scoring recipe

```python
def dice_score(pred, gt):
    intersection = np.sum(pred * gt)
    union = np.sum(pred) + np.sum(gt)
    if union == 0:
        return 1.0
    return 2.0 * intersection / union
```

## Common pitfalls

- LGG enhancing tumor segmentation yields 0.00 DSC due to intact blood-brain barrier and lack of contrast enhancement in most LGG cases.
- Cross-validation on training data may overestimate performance compared to an independent test set, as noted by the authors.
- Performance varies significantly between HGG and LGG cohorts due to boundary ambiguity and dataset size differences.

## Evidence (verbatim from paper)

> Table 3 tabulates the DSC results of our cross-validated segmentation results for the HGG and LGG cases, respectively. ... our method achieved 0.81 DSC for the enhancing tumor segmentation in the HGG cohort. By using our method, the enhancing tumor segmentation for the LGG cases by using the T1c images only is not successful.

## Citation

```bibtex
@misc{dong2017automatic,
  title={Automatic Brain Tumor Detection and Segmentation Using U-Net Based Fully Convolutional Networks},
  author={Dong et al. (2017)},
  year={2017},
  note={arXiv:1705.03820}
}
```

- arXiv: 1705.03820