Optimam Mammography Cad Eval

optimam-mammography-cad-eval

Large-scale mammography CAD with Deformable Conv-Nets — Morrell et al. (2019) (arXiv:1902.07323, 2019)

What this evaluates

Evaluates a deep learning object detection and classification system for identifying malignant lesions in mammographic images. It probes the model's ability to handle high-resolution medical imaging data and diverse lesion morphologies (e.g., microcalcifications, masses) using deformable convolutions.

Datasets

• Optimam — total ?; splits: train (-1)

Metrics

• AUC (primary) — range: [0, 1]
  • Area Under the Receiver Operating Characteristic Curve, measuring the trade-off between true positive rate and false positive rate across classification thresholds.

Input / output format

Input: Whole mammographic images (CC and MLO views) at high resolution (~2545x2545 pixels), potentially augmented via rotations and flips.

Output: Probability of malignancy lesion per image/view, aggregated across views and augmentations by taking the mean per laterality and then the maximum over both sides.

Scoring recipe

fpr, tpr, _ = roc_curve(gold, predictions)
auc = auc(fpr, tpr)
return auc

Common pitfalls

• Aggregation rule is specific: mean over views/augmentations per laterality, then max over sides, not a simple global average.
• High-resolution input (~2545x2545) requires careful memory management and downsampling, affecting feature extraction compared to standard natural image CNNs.

Evidence (verbatim from paper)

Each subjects’s probability of malignancy lesion was calculated by computing the mean over views and augmentations for each laterality, and then taking the maximum over the two sides... enabling superior detection performance on the Optimam dataset with an AUC of 0.879 in the DREAMS challenge

Citation

@misc{morrell2019largescale,
  title={Large-scale mammography CAD with Deformable Conv-Nets},
  author={Morrell et al. (2019)},
  year={2019},
  note={arXiv:1902.07323}
}

• arXiv: 1902.07323