Multi Oscc Eval

multi-oscc-eval

A High Magnifications Histopathology Image Dataset for Oral Squamous Cell Carcinoma Diagnosis and Prognosis — Guan et al. (2025) (arXiv:2507.16360, 2025)

What this evaluates

Evaluates vision models on diagnostic and prognostic classification of Oral Squamous Cell Carcinoma using high-magnification histopathology images. It probes multi-task learning capabilities, feature fusion across multiple tissue slices, and the impact of stain normalization and image resolution on clinical prediction accuracy.

Datasets

• Multi-OSCC — total 1325; splits: train (-1), val (-1), test (-1)

Metrics

• AUC (primary) — range: [0, 1]
  • Area under the receiver operating characteristic curve. Calculated with 95% confidence intervals via bootstrap estimation.
• accuracy — range: [0, 1]
  • Proportion of correctly classified instances.
• F1 score — range: [0, 1]
  • Harmonic mean of precision and recall.
• recall — range: [0, 1]
  • True positive rate.
• precision — range: [0, 1]
  • Positive predictive value.

Input / output format

Input: High-magnification histopathology images (×200, ×400, ×1000) extracted from core and edge tumor regions, typically processed at 512×512 resolution.

Output: Binary classification predictions for six clinical tasks: tumor recurrence (REC), lymph node metastasis (LNM), tumor differentiation (TD), tumor invasion (TI), cancer embolus (CE), and perineural invasion (PI).

Scoring recipe

def compute_auc(y_true, y_pred):
    fpr, tpr, _ = roc_curve(y_true, y_pred)
    auc_val = auc(fpr, tpr)
    boot_aucs = []
    for _ in range(1000):
        idx = np.random.choice(len(y_true), len(y_true), replace=True)
        boot_aucs.append(auc(fpr[idx], tpr[idx]))
    ci = (np.percentile(boot_aucs, 2.5), np.percentile(boot_aucs, 97.5))
    return auc_val, ci

Common pitfalls

• Stain normalization significantly degrades performance on the recurrence prediction (REC) task while improving others.
• Multi-task learning reduces average AUC by 3.34% due to task imbalance, making single-task benchmarks more reliable for individual predictions.
• Downsampling full-resolution images (2592×1944) to 512×512 is necessary for memory but loses fine-grained histological details.

Evidence (verbatim from paper)

For evaluation, we use five metrics: accuracy (Acc), area under the receiver operating characteristic curve (AUC), F1 score, recall, and precision. Although the final benchmark reports all metrics, the AUC is the primary metric used to select the best-performing model during the analysis phase.

Citation

@misc{guan2025highmagnifications,
  title={A High Magnifications Histopathology Image Dataset for Oral Squamous Cell Carcinoma Diagnosis and Prognosis},
  author={Guan et al. (2025)},
  year={2025},
  note={arXiv:2507.16360}
}

• arXiv: 2507.16360