Breast Cancer Screening Eval

breast-cancer-screening-eval

Multi-View Hypercomplex Learning for Breast Cancer Screening — Lopez et al. (2022) (arXiv:2204.05798, 2022)

What this evaluates

Evaluates multi-view deep learning models for medical image classification and segmentation, specifically testing the ability to model correlations between paired or multi-view medical images (mammograms, chest X-rays, brain MRIs) for tasks like lesion classification and survival prediction.

Datasets

• CBIS-DDSM — total 2478; splits: train (-1), test (-1)
• INbreast — total 410; splits: train (-1), test (-1)
• CheXpert — total 224316; splits: train (-1), test (-1)
• BraTS19 — total 210; splits: train (-1), test (-1)

Metrics

• AUC (primary) — range: [0, 1]
  • Area Under the Receiver Operating Characteristic Curve. Computed by plotting True Positive Rate against False Positive Rate across all classification thresholds and calculating the area under the resulting curve.
• classification accuracy — range: [0, 1]
  • The proportion of correctly classified instances out of the total number of instances, typically using a 0.5 probability threshold for binary decisions.
• Dice score — range: [0, 1]
  • Measures pixel-wise agreement between predicted and ground truth masks: Dice = 2 * |intersection| / (|mask_true| + |mask_pred|). Ranges from 0 (no overlap) to 1 (perfect overlap).

Input / output format

Input: Multi-view medical images (e.g., two or four mammogram views, frontal/lateral chest X-rays, or multi-modal brain MRI volumes) resized to task-specific dimensions (e.g., 600×500, 320×320, 128×128×128), optionally augmented with rotations, flips, or scaling.

Output: Binary classification probabilities or labels (benign/malignant, positive/negative, or survival prediction), or pixel-wise segmentation masks.

Scoring recipe

def compute_auc(y_true, y_pred):
    fpr, tpr, _ = roc_curve(y_true, y_pred)
    return auc(fpr, tpr)

def compute_accuracy(y_true, y_pred):
    preds = (y_pred >= 0.5).astype(int)
    return np.mean(y_true == preds)

def compute_dice(y_true, y_pred):
    intersection = np.sum(y_true * y_pred)
    return (2.0 * intersection) / (np.sum(y_true) + np.sum(y_pred))

Common pitfalls

• CBIS-DDSM contains only biopsy-proven positive cases (benign or malignant) with no healthy controls, which may overestimate real-world screening performance.
• INbreast uses BI-RADS labels rather than pathological ground truth; categories 4, 5, and 6 are mapped to positive and 1, 2 to negative, while category 3 is excluded, requiring careful label handling.
• Four-view experiments are restricted to INbreast only, as CBIS-DDSM lacks sufficient full-exam cases to create non-leaking four-view splits.

Evidence (verbatim from paper)

We adopt AUC (Area Under the ROC Curve) as the main performance metric to evaluate our models, as it is one of the most common metrics employed in medical imaging tasks. The ROC curve summarizes the trade-off between True Positive Rate (TPR) and False Positive Rate (FPR) for a predictive model using different probability thresholds. To further assess network performance, we additionally evaluate our models in terms of classification accuracy. Finally, we employ the Dice score for the segmentation task, which measures the pixel-wise agreement between a predicted mask and its corresponding ground truth.

Citation

@misc{lopez2022multiview,
  title={Multi-View Hypercomplex Learning for Breast Cancer Screening},
  author={Lopez et al. (2022)},
  year={2022},
  note={arXiv:2204.05798}
}

• arXiv: 2204.05798