Medical Segmentation Eval

---
name: medical-segmentation-eval
description: This evaluation benchmarks 2D medical image segmentation models across three diverse clinical datasets (polyp, skin lesion, cardiac ultrasound). It probes the cross-domain transferability and segmentation accuracy of general-purpose vision models versus specialized medical architectures. Use when the user wants to benchmark on NeoPolyp, CAMUS, ISIC'18, or asks about evaluating this task. Reports mDSC.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2603.13044
  bibtex_key: [REDACTED]
  confidence: high
---

# medical-segmentation-eval

> Are General-Purpose Vision Models All We Need for 2D Medical Image Segmentation? A Cross-Dataset Empirical Study — Borst et al. (2026) (arXiv:2603.13044, 2026)

## What this evaluates

This evaluation benchmarks 2D medical image segmentation models across three diverse clinical datasets (polyp, skin lesion, cardiac ultrasound). It probes the cross-domain transferability and segmentation accuracy of general-purpose vision models versus specialized medical architectures.

## Datasets

- **NeoPolyp** — total ?; splits: 5-fold CV (-1)
- **CAMUS** — total ?; splits: 5-fold CV (-1)
- **ISIC'18** — total ?; splits: 5-fold CV (-1)

## Metrics

- `mDSC` **(primary)** — range: percent
  - Mean Dice Similarity Coefficient. Calculated as the average Dice score across all segmentation classes, where Dice = 2 * |prediction ∩ ground_truth| / (|prediction| + |ground_truth|).
- `mIoU` — range: percent
  - Mean Intersection over Union. Average IoU across all classes, where IoU = |prediction ∩ ground_truth| / |prediction ∪ ground_truth|.
- `mRec` — range: percent
  - Mean Recall. Average recall across all segmentation classes.
- `mPrec` — range: percent
  - Mean Precision. Average precision across all segmentation classes.

## Input / output format

**Input**: 2D medical images (RGB endoscopic/skin or grayscale ultrasound) with corresponding pixel-level segmentation masks.

**Output**: Pixel-wise segmentation masks matching input dimensions, with class labels per dataset annotation scheme.

## Scoring recipe

```python
def compute_mDSC(pred_mask, gold_mask, num_classes):
    dice_scores = []
    for c in range(num_classes):
        pred_c = (pred_mask == c)
        gold_c = (gold_mask == c)
        intersection = np.sum(pred_c & gold_c)
        union = np.sum(pred_c | gold_c)
        dice = (2.0 * intersection) / (union + 1e-6)
        dice_scores.append(dice)
    return np.mean(dice_scores) * 100  # Paper reports as percentage
```

## Common pitfalls

- Performance varies significantly across classes; e.g., non-neoplastic polyps (C1) are consistently harder to segment than neoplastic ones.
- Evaluations use 5-fold cross-validation rather than a fixed held-out test set, so results should be interpreted as averaged fold metrics.
- Cross-dataset performance gaps are not uniform; GP-VMs show the largest advantage on NeoPolyp but only marginal gains on ISIC'18 and CAMUS.

## Evidence (verbatim from paper)

> Table 3 reports the 5-fold CV results, using mDSC as main performance metric. Measured by the average mDSC across all three datasets, the top-performing models are exclusively GP-VMs: VW-MiT (91.0%), VW-Conv and TransNeXt (both 90.9%), followed by InternImage (90.8%) as well as SegNeXt and SegFormer (both 90.7%).

## Citation

```bibtex
@misc{borst2026generalpurpose,
  title={Are General-Purpose Vision Models All We Need for 2D Medical Image Segmentation? A Cross-Dataset Empirical Study},
  author={Borst et al. (2026)},
  year={2026},
  note={arXiv:2603.13044}
}
```

- arXiv: 2603.13044