Isles24 Segmentation Eval

---
name: isles24-segmentation-eval
description: Evaluates the ability of models to perform 3D medical image segmentation for stroke lesion (infarct) and vessel occlusion detection using longitudinal multimodal CT and MRI scans. Use when the user wants to benchmark on ISLES'24, or asks about evaluating this task. Reports Dice Similarity Coefficient (DSC).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2408.11142
  bibtex_key: [REDACTED]
  confidence: high
---

# isles24-segmentation-eval

> ISLES'24 -- A Real-World Longitudinal Multimodal Stroke Dataset — Riedel et al. (2024) (arXiv:2408.11142, 2024)

## What this evaluates

Evaluates the ability of models to perform 3D medical image segmentation for stroke lesion (infarct) and vessel occlusion detection using longitudinal multimodal CT and MRI scans.

## Datasets

- **ISLES'24** — total 245; splits: train (149), test (96); repo https://github.com/ezequieldrosa/isles24

## Metrics

- `Dice Similarity Coefficient (DSC)` **(primary)** — range: [0, 1]
  - DSC = 2 * |Prediction ∩ Ground Truth| / (|Prediction| + |Ground Truth|). Standard metric for binary segmentation challenges in medical imaging.

## Input / output format

**Input**: 3D NIfTI medical images (NCCT, CTA, CTP, DWI, ADC, FLAIR) co-registered to NCCT space, plus tabular clinical data.

**Output**: 3D binary segmentation masks (NIfTI) indicating infarct lesions and/or vessel occlusion locations.

## 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 (2.0 * intersection) / union if union > 0 else 0.0
```

## Common pitfalls

- Test set is hidden and evaluated via the ISLES'24 Grand Challenge website, not locally.
- Images require strict co-registration to NCCT space; misalignment severely impacts segmentation metrics.
- Clinical tabular data is randomly altered by ±5% for anonymization, so models should not rely on exact numerical values for clinical variables.

## Evidence (verbatim from paper)

> The loss was a combined CE, Dice, Skeleton Recall 37 , and a customized topology loss.

## Citation

```bibtex
@misc{riedel2024isles24,
  title={ISLES'24 -- A Real-World Longitudinal Multimodal Stroke Dataset},
  author={Riedel et al. (2024)},
  year={2024},
  note={arXiv:2408.11142}
}
```

- arXiv: 2408.11142