Mammography Abnormality Eval

---
name: mammography-abnormality-eval
description: Evaluates deep CNN architectures for classifying mammographic abnormalities (calcifications and masses) and localizing them using class activation maps. It probes the model's ability to learn patch-based features and generalize to full-image localization without explicit spatial supervision. Use when the user wants to benchmark on Mammography dataset (unspecified), or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1803.01906
  bibtex_key: [REDACTED]
  confidence: medium
---

# mammography-abnormality-eval

> Abnormality Detection in Mammography using Deep Convolutional Neural Networks — Xi et al. (2018) (arXiv:1803.01906, 2018)

## What this evaluates

Evaluates deep CNN architectures for classifying mammographic abnormalities (calcifications and masses) and localizing them using class activation maps. It probes the model's ability to learn patch-based features and generalize to full-image localization without explicit spatial supervision.

## Datasets

- **Mammography dataset (unspecified)** — total ?; splits: train (-1), test (-1)

## Metrics

- `accuracy` **(primary)** — range: percent
  - Percentage of correctly classified image patches out of the total number of patches in a fold. Computed as (True Positives + True Negatives) / Total.

## Input / output format

**Input**: Cropped image patches for classification; full mammogram images for localization.

**Output**: Classification label (calcification or mass) for patches; class activation map (heatmap) for full images.

## Scoring recipe

```python
def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100
```

## Common pitfalls

- Localization is evaluated qualitatively via visual inspection of CAM heatmaps against ground-truth masks, not with quantitative metrics like IoU or Dice.
- The paper uses 5-fold cross-validation on patches but does not specify if splits are patient-level or patch-level, risking data leakage in medical imaging evaluations.

## Evidence (verbatim from paper)

> Running cross-validation on training and testing datasets and computing mean accuracies across the five folds give the final accuracy results in Table [II](#S4.T2 "TABLE II ‣ IV-A Comparison of Different Deep CNN Architectures ‣ IV Results ‣ Abnormality Detection in Mammography using Deep Convolutional Neural Networks"). VGGNet achieves the highest accuracy for classifying calcifications and GoogleNet receives the best performance for classifying masses. The highest overall accuracy is also achieved by VGGNet at 92.53%.

## Citation

```bibtex
@misc{xi2018abnormality,
  title={Abnormality Detection in Mammography using Deep Convolutional Neural Networks},
  author={Xi et al. (2018)},
  year={2018},
  note={arXiv:1803.01906}
}
```

- arXiv: 1803.01906