Mammographic Lesion Seg Eval

---
name: mammographic-lesion-seg-eval
description: Evaluates lightweight CNN architectures for pixel-wise lesion segmentation in mammograms. It measures segmentation accuracy and computational efficiency, while also probing cross-dataset generalization under domain shift and the sensitivity of performance metrics to post-processing thresholds. Use when the user wants to benchmark on INbreast, DMID, or asks about evaluating this task. Reports Dice Score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.23899
  bibtex_key: oliveira2026mammographic
  confidence: high
---

# mammographic-lesion-seg-eval

> Mammographic Lesion Segmentation with Lightweight Models: A Comparative Study — Oliveira et al. (2026) (arXiv:2604.23899, 2026)

## What this evaluates

Evaluates lightweight CNN architectures for pixel-wise lesion segmentation in mammograms. It measures segmentation accuracy and computational efficiency, while also probing cross-dataset generalization under domain shift and the sensitivity of performance metrics to post-processing thresholds.

## Datasets

- **INbreast** — total ?; splits: 5-fold CV (-1)
- **DMID** — total ?; splits: test (-1)

## Metrics

- `Dice Score` **(primary)** — range: [0, 1]
  - Dice coefficient measures the overlap between predicted and ground-truth masks: 2 * |A ∩ B| / (|A| + |B|). It is the primary metric for model selection.
- `IoU` — range: [0, 1]
  - Intersection over Union calculates the ratio of the intersection area to the union area of predicted and ground-truth masks.
- `Recall` — range: [0, 1]
  - Recall (True Positive Rate) measures the proportion of actual lesion pixels correctly identified by the model.

## Input / output format

**Input**: Grayscale mammographic images from the INbreast or DMID datasets.

**Output**: Binary segmentation masks (probability maps thresholded at a specified value, e.g., 0.5).

## Scoring recipe

```python
def compute_metrics(pred_mask, gt_mask):
    intersection = np.logical_and(pred_mask, gt_mask).sum()
    union = np.logical_or(pred_mask, gt_mask).sum()
    gt_sum = gt_mask.sum()
    dice = 2 * intersection / (pred_mask.sum() + gt_sum) if (pred_mask.sum() + gt_sum) > 0 else 0
    iou = intersection / union if union > 0 else 0
    recall = intersection / gt_sum if gt_sum > 0 else 0
    return dice, iou, recall
```

## Common pitfalls

- Domain shift primarily degrades precision and boundary delineation (lower Dice/IoU) while preserving recall, which can mask generalization failures if only recall is monitored.
- Threshold selection heavily influences Dice and IoU scores but has minimal impact on Recall, requiring explicit threshold reporting for fair comparison.
- Pairwise statistical tests (Wilcoxon with Bonferroni correction) showed no significant differences between top lightweight models despite numerical gaps, warning against overinterpreting small metric deltas.

## Evidence (verbatim from paper)

> The Table 4 shows the Dice score, IoU, and Recall considering a segmentation threshold of 0.5.

## Citation

```bibtex
@misc{oliveira2026mammographic,
  title={Mammographic Lesion Segmentation with Lightweight Models: A Comparative Study},
  author={Oliveira et al. (2026)},
  year={2026},
  note={arXiv:2604.23899}
}
```

- arXiv: 2604.23899