Trackrad2025 Eval

---
name: trackrad2025-eval
description: Probes the capability of real-time tumor and surrogate localization in MRI-guided radiotherapy using 2D sagittal cine MRI sequences. It evaluates how well algorithms can track anatomical motion across varying frame rates and multi-vendor MRI-linac hardware under clinically relevant conditions. Use when the user wants to benchmark on TrackRAD2025, or asks about evaluating this task. Reports tracking performance.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2503.19119
  bibtex_key: wang2025trackrad2025
  confidence: medium
---

# trackrad2025-eval

> TrackRAD2025 challenge dataset: Real-time tumor tracking for MRI-guided radiotherapy — Wang et al. (2025) (arXiv:2503.19119, 2025)

## What this evaluates

Probes the capability of real-time tumor and surrogate localization in MRI-guided radiotherapy using 2D sagittal cine MRI sequences. It evaluates how well algorithms can track anatomical motion across varying frame rates and multi-vendor MRI-linac hardware under clinically relevant conditions.

## Datasets

- **TrackRAD2025** — total 585; splits: train (527), test (58); repo https://doi.org/10.57967/hf/4539

## Metrics

- `tracking performance` **(primary)** — range: percent
  - Evaluates the accuracy of real-time tumor or surrogate localization across time-resolved 2D cine MRI frames. The exact mathematical formula is not specified in the provided text, but the challenge focuses on online, time-resolved tracking accuracy under clinically relevant motion conditions.

## Input / output format

**Input**: 2D sagittal cine MRI sequences (20–20543 frames per scan) acquired at 1–8 Hz from 0.35 T and 1.5 T MRI-linac scanners.

**Output**: Predicted tumor or surrogate segmentations (or tracking coordinates) for each temporal frame in the input sequence.

## Scoring recipe

```python
def compute_tracking_performance(predictions, ground_truth):
    # predictions and ground_truth are lists of binary masks or coordinates per frame
    scores = []
    for pred, gt in zip(predictions, ground_truth):
        # Standard segmentation overlap or displacement error calculation
        intersection = np.sum(pred & gt)
        union = np.sum(pred | gt)
        score = (2 * intersection) / (union + 1e-6)
        scores.append(score)
    return np.mean(scores) * 100
```

## Common pitfalls

- Failing to account for variable frame rates (1–8 Hz) and temporal resolution differences across multi-vendor MRI-linac data.
- Incorrectly treating the 477 unlabeled training frames as ground truth, which are explicitly marked as unlabeled in the challenge protocol.
- Evaluating offline post-processing instead of real-time/online tracking performance, which is the core clinical requirement.

## Evidence (verbatim from paper)

> The TrackRAD2025 challenge provides a standardized platform for comparing online, time-resolved 2D cine-MRI-based tracking performance, addressing the lack of public benchmarks for real-time motion localization in MRI-linac workflows.

## Citation

```bibtex
@misc{wang2025trackrad2025,
  title={TrackRAD2025 challenge dataset: Real-time tumor tracking for MRI-guided radiotherapy},
  author={Wang et al. (2025)},
  year={2025},
  note={arXiv:2503.19119}
}
```

- arXiv: 2503.19119