Ct Brain Segmentation Eval

---
name: ct-brain-segmentation-eval
description: Evaluates the ability of segmentation models to accurately delineate brain tissue, cerebrospinal fluid (CSF), and subdural hematomas in post-operative CT scans of hydrocephalic infants. It probes robustness to intensity overlap, anatomical distortion, and limited training data in a real-world clinical setting. Use when the user wants to benchmark on CURE Children's Hospital of Uganda CT Brain Dataset, or asks about evaluating this task. Reports dice-overlap coefficient.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1712.03993
  bibtex_key: [REDACTED]
  confidence: high
---

# ct-brain-segmentation-eval

> Learning Based Segmentation of CT Brain Images: Application to Post-Operative Hydrocephalic Scans — Cherukuri et al. (2017) (arXiv:1712.03993, 2017)

## What this evaluates

Evaluates the ability of segmentation models to accurately delineate brain tissue, cerebrospinal fluid (CSF), and subdural hematomas in post-operative CT scans of hydrocephalic infants. It probes robustness to intensity overlap, anatomical distortion, and limited training data in a real-world clinical setting.

## Datasets

- **CURE Children's Hospital of Uganda CT Brain Dataset** — total 32; splits: train (-1), test (15)

## Metrics

- `dice-overlap coefficient` **(primary)** — range: [0, 1]
  - Computed per class as DO(A,B) = 2|A∩B|/(|A|+|B|), where A and B are the predicted and ground truth masks. Evaluates to 1 only when masks are identical. Results are averaged across all test patients.

## Input / output format

**Input**: Stack of 2D CT slices (512x512 pixels, 3-10mm thickness) per patient, processed using square patches (e.g., 11x11 or 13x13).

**Output**: Per-pixel class labels for three segments: Brain, CSF, and Subdural hematoma.

## Scoring recipe

```python
def compute_dice(pred_mask, gt_mask):
    intersection = np.sum(pred_mask & gt_mask)
    union = np.sum(pred_mask) + np.sum(gt_mask)
    return 2 * intersection / union if union > 0 else 0.0

dice_scores = []
for patient in test_patients:
    for class_label in ['Brain', 'CSF', 'Subdural']:
        pred = get_prediction(patient, class_label)
        gt = get_ground_truth(patient, class_label)
        dice_scores.append(compute_dice(pred, gt))
mean_dice = np.mean(dice_scores)
```

## Common pitfalls

- Averaging is performed over patients rather than over pixels or patches, so small patient counts can significantly skew results.
- The metric is computed separately for each of the three classes (Brain, CSF, Subdural) rather than as a single global Dice score.
- Slice thickness varies (3-10mm), which affects 3D volume interpretation but the evaluation protocol applies the metric per slice/patch.

## Evidence (verbatim from paper)

> To validate our results, we used the dice-overlap coefficient, which for regions A and B is defined as DO(A,B) = 2|A∩B|/(|A|+|B|). Note, DO(A,B) evaluates to 1, only when A=B. The dice-overlap is computed for each method by using carefully obtained manually segmented results under the supervision of an expert neurosurgeon - (SJS).

## Citation

```bibtex
@misc{cherukuri2017segmentation,
  title={Learning Based Segmentation of CT Brain Images: Application to Post-Operative Hydrocephalic Scans},
  author={Cherukuri et al. (2017)},
  year={2017},
  note={arXiv:1712.03993}
}
```

- arXiv: 1712.03993