Topoc Cancer Diagnosis Eval

topoc-cancer-diagnosis-eval

TopOC: Topological Deep Learning for Ovarian and Breast Cancer Diagnosis — Fatema et al. (2024) (arXiv:2410.09818, 2024)

What this evaluates

Evaluates histopathology image classification for ovarian and breast cancer diagnosis using topological deep learning features combined with CNNs. Probes the model's ability to differentiate cancer subtypes and benign/malignant cases from microscopic tissue tiles.

Datasets

• UBC-OCEAN — total 34285; splits: train (31203), test (3082)
• BREAKHIS — total 7909; splits: train (5536), test (2373)

Metrics

• Balanced Accuracy (primary) — range: [0, 1]
  • Average of recall obtained on each class. Computed as (1/C) * sum(TP_c / (TP_c + FN_c)) for C classes.
• Accuracy — range: [0, 1]
  • Fraction of correctly classified samples out of the total number of samples.
• AUC — range: [0, 1]
  • Area under the Receiver Operating Characteristic (ROC) curve, measuring the trade-off between true positive and false positive rates.
• Sensitivity — range: [0, 1]
  • True positive rate: TP / (TP + FN).
• Specificity — range: [0, 1]
  • True negative rate: TN / (TN + FP).

Input / output format

Input: 2D histopathology image tiles (224×224 for UBC-OCEAN, 700×460 for BREAKHIS) processed through RGB color channels to extract topological persistence vectors, concatenated with CNN feature embeddings.

Output: Class predictions: 5-class softmax output for UBC-OCEAN (clear cell, endometrioid, high-grade serous, low-grade serous, mucinous carcinoma) or 2-class output for BREAKHIS (Benign, Malignant).

Scoring recipe

def compute_metrics(y_true, y_pred, n_classes):
    acc = (y_true == y_pred).mean()
    bal_acc = 0.0
    for c in range(n_classes):
        tp = ((y_true == c) & (y_pred == c)).sum()
        fn = ((y_true == c) & (y_pred != c)).sum()
        bal_acc += tp / (tp + fn) if (tp + fn) > 0 else 0.0
    bal_acc /= n_classes
    return acc, bal_acc

Common pitfalls

• Class imbalance in UBC-OCEAN makes standard accuracy misleading; balanced accuracy is required for fair comparison.
• BREAKHIS results vary significantly by magnification factor (40x–400x); metrics must be reported per magnification level.
• Topological vector dimensionality (50–400) and CNN backbone choice heavily influence performance; results are not directly comparable across different feature configurations.

Evidence (verbatim from paper)

We used the pre-defined training and test split (91:9) provided by the challenge, i.e., 31,203 training and 3,082 test images. Our second dataset, BREAKHIS, is the Breast Cancer Histopathology dataset, consisting of 7909 samples with two classes, Benign and Malignant. For this dataset, we used a commonly employed 70:30 training-test split. The performance metric was balanced accuracy, and the winning entry achieved 66% balanced accuracy for this five-class classification task.

Citation

@misc{fatema2024topoc,
  title={TopOC: Topological Deep Learning for Ovarian and Breast Cancer Diagnosis},
  author={Fatema et al. (2024)},
  year={2024},
  note={arXiv:2410.09818}
}

• arXiv: 2410.09818