Nyu Breast Cancer Seg Eval

nyu-breast-cancer-seg-eval

Weakly-supervised High-resolution Segmentation of Mammography Images for Breast Cancer Diagnosis — Liu et al. (2021) (arXiv:2106.07049, 2021)

What this evaluates

Evaluates weakly-supervised segmentation and classification performance on high-resolution mammography images for detecting malignant and benign breast lesions.

Datasets

• NYU Breast Cancer Screening Dataset v1.0 — total 229426; splits: train (186816), val (28462), test (14148)

Metrics

• Dice similarity coefficient (primary) — range: [0, 1]
  • 2 * |A ∩ B| / (|A| + |B|), measuring overlap between predicted and ground-truth lesion masks.
• Pixel Average Precision (PxAP) — range: [0, 1]
  • Average area under the precision-recall curve computed per pixel. Thresholds are either chosen per image (image-level) or fixed across the test set (dataset-level).
• Classification AUC — range: [0, 1]
  • Area under the Receiver Operating Characteristic curve for binary classification of malignant or benign lesion presence.

Input / output format

Input: High-resolution mammography images (2944×1920 px) per standard view (R-CC, L-CC, R-MLO, L-MLO), with exam-level binary labels indicating presence of malignant or benign lesions.

Output: Per-pixel saliency maps or segmentation masks for malignant and benign lesions, plus binary classification scores for lesion presence.

Scoring recipe

# Dice coefficient
intersection = np.sum(pred_mask & gt_mask)
dice = 2.0 * intersection / (np.sum(pred_mask) + np.sum(gt_mask))

# Pixel Average Precision (PxAP)
ap_scores = []
for pixel_idx in range(num_pixels):
    pred_scores = pred_saliency[:, pixel_idx]
    gt_labels = gt_mask[:, pixel_idx]
    ap_scores.append(compute_auc_pr(pred_scores, gt_labels))
pxap = np.mean(ap_scores)

Common pitfalls

• Dataset is highly imbalanced (~99% exams have no lesions); models require balanced sampling during training to avoid trivial solutions.
• PxAP results vary significantly depending on whether the threshold is optimized per image (image-level) or fixed globally (dataset-level).
• Local module segmentation can fail entirely if the patch-selection step misses the lesion or if the lesion exceeds the patch size.

Evidence (verbatim from paper)

To measure classification performance, we report the area under the ROC curve (AUC) for identifying breasts with both malignant and benign lesions. To evaluate localization ability, we use the Dice similarity coefficient and pixel average precision (PxAP) (Choe et al., 2020). PxAP is the average of the area under the precision-recall curve for each pixel. The threshold to compute the precision and recall is either chosen for each image (image-level PxAP) or fixed for the whole test set (dataset-level PxAP).

Citation

@misc{liu2021glam,
  title={Weakly-supervised High-resolution Segmentation of Mammography Images for Breast Cancer Diagnosis},
  author={Liu et al. (2021)},
  year={2021},
  note={arXiv:2106.07049}
}

• arXiv: 2106.07049