Wsi Classification Eval

wsi-classification-eval

Promptable Representation Distribution Learning and Data Augmentation for Gigapixel Histopathology WSI Analysis — Tang et al. (2024) (arXiv:2412.14473, 2024)

What this evaluates

Evaluates whole-slide image (WSI) classification performance using self-supervised patch representations and feature-space data augmentation. It probes how well distribution-guided representation learning captures discriminative histopathological patterns for diagnostic subtyping.

Datasets

• USTC-EGFR — total 754; splits: train (-1), val (-1), test (-1)
• TCGA-EGFR — total 696; splits: train (-1), val (-1), test (-1)
• TCGA-LUNG-3K — total 3064; splits: train (-1), val (-1), test (-1)

Metrics

• average accuracy — range: [0, 1]
  • Proportion of correctly classified WSIs out of the total number of WSIs in the evaluation set.
• micro-average area under the curve (AUC) (primary) — range: [0, 1]
  • Area under the Receiver Operating Characteristic (ROC) curve computed by pooling predictions across all classes (micro-averaging) before calculating the AUC.
• macro-average F1 score — range: [0, 1]
  • Harmonic mean of precision and recall calculated per class, then averaged across all classes (macro-averaging) to treat all classes equally regardless of frequency.

Input / output format

Input: WSIs segmented into non-overlapping 224×224 patches at 20× magnification, encoded via a ViT-S/16 backbone, and aggregated using Multiple Instance Learning (MIL) backbones (CLAM, TransMIL, or DTFD-MIL).

Output: Predicted WSI class label (e.g., EGFR mutation status or lung cancer subtype).

Scoring recipe

def compute_metrics(y_true, y_pred, y_prob):
    acc = np.mean(y_true == y_pred)
    auc = roc_auc_score(y_true, y_prob, average='micro', multi_class='ovr')
    f1 = f1_score(y_true, y_pred, average='macro')
    return {'accuracy': acc, 'micro-AUC': auc, 'macro-F1': f1}

Common pitfalls

• Splits are performed at the patient level (6:1:3 ratio) to prevent data leakage, not at the WSI level.
• Results are benchmarked across three different MIL aggregation backbones (CLAM, TransMIL, DTFD-MIL), so performance varies significantly by architecture.
• Feature-space augmentation is evaluated rather than image-space, meaning metrics reflect representation discriminability rather than raw pixel-level classification.

Evidence (verbatim from paper)

Our evaluation metrics include average accuracy, micro-average area under the curve (AUC), and macro-average F1 score.

Citation

@misc{tang2024prdl,
  title={Promptable Representation Distribution Learning and Data Augmentation for Gigapixel Histopathology WSI Analysis},
  author={Tang et al. (2024)},
  year={2024},
  note={arXiv:2412.14473}
}

• arXiv: 2412.14473