Breast Cancer Mammography Eval

breast-cancer-mammography-eval

Deep Learning to Improve Breast Cancer Early Detection on Screening Mammography — Shen et al. (2017) (arXiv:1708.09427, 2017)

What this evaluates

Evaluates deep learning models for binary classification of breast cancer (malignant vs. benign/normal) on screening mammograms. It probes the model's ability to generalize across different mammography platforms (film vs. digital) and transfer learned features from patch-level to whole-image classification without requiring costly lesion-level annotations.

Datasets

• CBIS-DDSM — total 2478; splits: train (1903), val (199), test (376)
• INbreast — total 410; splits: train (-1), val (-1)

Metrics

• AUC (primary) — range: [0, 1]
  • Area under the Receiver Operating Characteristic (ROC) curve, computed by integrating the true positive rate against the false positive rate across all classification thresholds.

Input / output format

Input: Mammogram images (converted to PNG, resized to 1152x896) or 224x224 patches sampled from ROIs/background. Labels: binary (malignant vs. benign/normal) for whole images, or 5-class (background, malignant/benign mass/calcification) for patches.

Output: Predicted probability score for malignancy per image. For ensembles, scores are averaged across four models after test-time augmentation (horizontal/vertical flips).

Scoring recipe

def compute_auc(y_true, y_pred):
    # y_true: binary labels (1=malignant, 0=benign/normal)
    # y_pred: predicted probabilities of malignancy
    fpr, tpr, _ = roc_curve(y_true, y_pred)
    return auc(fpr, tpr)

Common pitfalls

• Splits must be performed at the patient level, not image level, to prevent data leakage from multiple views of the same patient.
• INbreast lacks pathological confirmation; labels are assigned based on BI-RADS categories (1-2 negative, 4-6 positive), excluding BI-RADS 3 cases.
• Patch sampling strategy (S1 vs S10) significantly impacts performance, as overlapping patches around ROIs provide crucial contextual information for whole-image classification.

Evidence (verbatim from paper)

The purpose of this study was to predict the malignant vs. benign (or normal) status of each image. We performed an 85-15 split on the patient-level data to create independent training and test sets. ... The models were evaluated by per-image AUCs on the test set.

Citation

@misc{shen2017deeplearning,
  title={Deep Learning to Improve Breast Cancer Early Detection on Screening Mammography},
  author={Shen et al. (2017)},
  year={2017},
  note={arXiv:1708.09427}
}

• arXiv: 1708.09427