Brats Tcga Tumor Seg Eval

---
name: brats-tcga-tumor-seg-eval
description: Evaluates the capability of deep learning models to segment brain tumors from multi-modal MRI scans. It probes volumetric overlap accuracy and boundary localization precision across distinct tumor sub-regions (enhancing tumor, tumor core, whole tumor). Use when the user wants to benchmark on BraTS-Glioma (BraTS 2020), TCGA LGG, or asks about evaluating this task. Reports Dice Similarity Coefficient (DSC).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2501.11196
  bibtex_key: behzadpour2025enhancing
  confidence: high
---

# brats-tcga-tumor-seg-eval

> Enhancing Brain Tumor Segmentation Using Channel Attention and Transfer learning — Majid Behzadpour et al. (arXiv:2501.11196, 2025)

## What this evaluates

Evaluates the capability of deep learning models to segment brain tumors from multi-modal MRI scans. It probes volumetric overlap accuracy and boundary localization precision across distinct tumor sub-regions (enhancing tumor, tumor core, whole tumor).

## Datasets

- **BraTS-Glioma (BraTS 2020)** — total 1251; splits: train (1100), test (151)
- **TCGA LGG** — total 3929; splits: train (-1), val (28), test (28)

## Metrics

- `Dice Similarity Coefficient (DSC)` **(primary)** — range: [0, 1]
  - Ratio of overlapping area between predicted and ground truth masks: DSC = 2|P ∩ G| / (|P| + |G|). Values closer to 1 indicate better volumetric overlap.
- `95th percentile Hausdorff Distance (HD95)` — range: other
  - Computes the 95th percentile of the maximum surface-to-surface distance between the predicted segmentation boundary and the ground truth boundary. Insensitive to extreme outliers by excluding the worst 5% of mismatches.

## Input / output format

**Input**: Multi-modal MRI volumes (T1, T1Gd, T2, T2-FLAIR for BraTS; unspecified modalities for TCGA LGG) paired with pixel-wise ground truth segmentation masks for tumor sub-regions.

**Output**: Predicted binary segmentation masks for each tumor sub-region (Enhancing Tumor, Tumor Core, Whole Tumor).

## 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):
    # Compute surface distances between predicted and ground truth boundaries
    surface_dists = compute_surface_distances(pred_mask, gt_mask)
    return np.percentile(surface_dists, 95)
```

## Common pitfalls

- DSC is highly sensitive to class imbalance; small tumor sub-regions (e.g., ET) often yield lower scores than larger ones (e.g., WT) even with good boundary alignment.
- HD95 is computed on boundary surfaces rather than volumes, so it can be disproportionately affected by small localized segmentation errors or image noise.
- The protocol requires separate evaluation for three distinct sub-regions (ET, TC, WT); reporting a single aggregate score without specifying the region is misleading.

## Evidence (verbatim from paper)

> to evaluate the performance of our brain tumor segmentation models, two common metrics were adopted: The Dice Similarity Coefficient-DSC, and the 95th percentile Hausdorff Distance-HD95, both widely used in medical image analysis. The Dice coefficient refers to the ratio of overlapping between the predicted tumor area and the ground truth annotation, it is defined as: $$ \mathrm {D S C} = \frac {2 | P \cap G |}{| P | + | G |} $$ where P is the predicted tumor region and G is the region of the ground truth. A higher Dice score indicates a better overlap between the real and predicted masks of tumors.

## Citation

```bibtex
@misc{behzadpour2025enhancing,
  title={Enhancing Brain Tumor Segmentation Using Channel Attention and Transfer learning},
  author={Majid Behzadpour et al.},
  year={2025},
  note={arXiv:2501.11196}
}
```

- arXiv: 2501.11196