# Inbreast Mammography Eval

> Probes the ability of deep learning models to detect malignancy in mammograms and generalize across different imaging scanners and patient populations. It specifically tests whether injecting stable, multi-scale topological features improves robustness to domain shifts compared to standard grayscale inputs. Use when the user wants to benchmark on INbreast, or asks about evaluating this task. Reports patient-level AUC.

- Skill: `qhjqhj00/inbreast-mammography-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/inbreast-mammography-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/inbreast-mammography-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/inbreast-mammography-eval

---


# inbreast-mammography-eval

> Topological Conditioning for Mammography Models via a Stable Wavelet-Persistence Vectorization — Fanning et al. (2025) (arXiv:2512.10151, 2025)

## What this evaluates

Probes the ability of deep learning models to detect malignancy in mammograms and generalize across different imaging scanners and patient populations. It specifically tests whether injecting stable, multi-scale topological features improves robustness to domain shifts compared to standard grayscale inputs.

## Datasets

- **INbreast** — total ?; splits: test (-1)

## Metrics

- `patient-level AUC` **(primary)** — range: [0, 1]
  - Area Under the Receiver Operating Characteristic curve computed at the patient level, aggregating predictions across all patches or proposals for a given patient.

## Input / output format

**Input**: Grayscale mammography images min-max normalized to [0,1], optionally concatenated with multi-scale wavelet-persistence channels. Images are embedded in a fixed grid and processed as overlapping 224×224 patches in a two-stage detection pipeline.

**Output**: Breast-level malignancy score (Stage I) or image-level detections and classification decisions (Stage II).

## Scoring recipe

```python
def compute_patient_auc(predictions, gold_labels):
    from sklearn.metrics import roc_auc_score
    return roc_auc_score(gold_labels, predictions)
```

## Common pitfalls

- The evaluation uses a deliberately limited training budget (625 iterations for Stage I, 100 for Stage II), which may not represent fully converged models and could bias performance comparisons.
- The protocol emphasizes domain shift robustness but does not specify exact scanner or population splits in the provided text, making reproducibility of the cross-domain evaluation difficult.
- Metrics are reported at the patient level rather than image or patch level, which aggregates multiple patch predictions and may obscure localization errors.

## Evidence (verbatim from paper)

> leading to a significant boost in patient-level AUC (from 0.55 to 0.75) on the INbreast cohort under a fixed training budget, demonstrating robustness to domain shifts and morphology sensitivity.

## Citation

```bibtex
@misc{fanning2025topological,
  title={Topological Conditioning for Mammography Models via a Stable Wavelet-Persistence Vectorization},
  author={Fanning et al. (2025)},
  year={2025},
  note={arXiv:2512.10151}
}
```

- arXiv: 2512.10151