Dyabd Segmentation Eval

---
name: dyabd-segmentation-eval
description: This benchmark evaluates the segmentation capabilities of deep learning models on dynamic abdominal MRI scans. It specifically probes how well models handle extreme anatomical variability caused by real-time muscle motion during breathing and Valsalva maneuvers, across few-shot, prompt-based, and fully automatic inference settings. Use when the user wants to benchmark on DyABD, or asks about evaluating this task. Reports Dice.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.23187
  bibtex_key: belton2026dyabd
  confidence: high
---

# dyabd-segmentation-eval

> DyABD: The Abdominal Muscle Segmentation in Dynamic MRI Benchmark — Belton et al. (2026) (arXiv:2604.23187, 2026)

## What this evaluates

This benchmark evaluates the segmentation capabilities of deep learning models on dynamic abdominal MRI scans. It specifically probes how well models handle extreme anatomical variability caused by real-time muscle motion during breathing and Valsalva maneuvers, across few-shot, prompt-based, and fully automatic inference settings.

## Datasets

- **DyABD** — total 311; splits: train (233), test (78); repo https://github.com/niamhbelton/DyABD-Segmentation

## Metrics

- `Dice` **(primary)** — range: [0, 1]
  - Computes the overlap between predicted and ground truth masks: 2 * |intersection| / (|prediction| + |ground_truth|). Scores are averaged across slices and muscle classes, then reported as the median per model.

## Input / output format

**Input**: 2D MRI slices (resized to 128×128, 256×256, or 1024×1024 depending on the model). For few-shot settings, a support set of N annotated slices is provided alongside query slices. For prompt-based settings, bounding box coordinates are provided per slice. Fully automatic settings receive only the raw slices.

**Output**: Binary segmentation mask for each of the four abdominal muscle classes per slice. Models perform inference per slice and per muscle, requiring four passes per volume to generate the final mask.

## Scoring recipe

```python
def dice_score(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 1.0
# Compute per slice, per muscle class, then average across all slices/classes
# Report median Dice across all test volumes
```

## Common pitfalls

- Inference is strictly 2D (slice-by-slice), ignoring 3D volumetric context despite the data being dynamic MRI volumes.
- Standard augmentations like flipping or rotating are explicitly disabled due to anatomical symmetry, deviating from typical medical imaging pipelines.
- Evaluation requires four separate model passes per volume (one per muscle class) to assemble the final segmentation mask.

## Evidence (verbatim from paper)

> The dataset evaluates supervised, few-shot, and zero-shot segmentation models, revealing that while existing methods achieve only a median Dice of 0.82, off-the-shelf zero-shot models (e.g., SAM, SAM2) outperform fine-tuned medical-specific models like MedSAM, highlighting a critical gap in generalizability and underscoring the need for robust zero-shot capabilities in medical image analysis under dynamic, low-data conditions.

## Citation

```bibtex
@misc{belton2026dyabd,
  title={DyABD: The Abdominal Muscle Segmentation in Dynamic MRI Benchmark},
  author={Belton et al. (2026)},
  year={2026},
  note={arXiv:2604.23187}
}
```

- arXiv: 2604.23187