Hepatobench Eval

hepatobench-eval

A Digital Pathology Resource for Liver Cancer Quantification with Datasets, Benchmarks, and Tools — Xiao et al. (2026) (arXiv:2604.22858, 2026)

What this evaluates

Evaluates pathology foundation models on fine-grained tissue classification of liver cancer patches and whole-slide tumor/non-tumor segmentation. It measures how well models can quantify tissue composition (e.g., fibrosis, necrosis, inflammation) within clinically defined regions to support reproducible digital pathology analysis.

Datasets

• HepatoBench — total ?; splits: train (-1), val (-1), test (-1)

Metrics

• F1-score (primary) — range: [0, 1]
  • Harmonic mean of precision and recall. Computed per class for the seven tissue categories and macro-averaged across classes for the final score.
• Dice coefficient (primary) — range: [0, 1]
  • 2 * |A ∩ B| / (|A| + |B|), where A and B are the predicted and ground truth segmentation masks. Measures spatial overlap between predicted and annotated tumor/non-tumor regions.
• Accuracy — range: [0, 1]
  • Proportion of correctly classified patches out of the total number of patches.
• IoU — range: [0, 1]
  • Intersection over Union: |A ∩ B| / |A ∪ B|, measuring the overlap ratio between predicted and ground truth segmentation masks.

Input / output format

Input: 150×150 pixel image patches at 20× magnification for classification; whole-slide images (WSIs) for segmentation.

Output: Per patch: one of seven tissue class labels (TUM, FIB, INF, NEC, NOR, REA, STE). Per WSI: a binary mask delineating tumor vs. non-tumor regions.

Scoring recipe

def compute_metrics(preds, gold):
    acc = sum(p == g for p, g in zip(preds, gold)) / len(gold)
    f1s = []
    for cls in ['TUM','FIB','INF','NEC','NOR','REA','STE']:
        tp = sum(1 for p,g in zip(preds,gold) if p==cls and g==cls)
        fp = sum(1 for p,g in zip(preds,gold) if p==cls and g!=cls)
        fn = sum(1 for p,g in zip(preds,gold) if p!=cls and g==cls)
        prec = tp/(tp+fp) if (tp+fp)>0 else 0
        rec = tp/(tp+fn) if (tp+fn)>0 else 0
        f1s.append(2*prec*rec/(prec+rec) if (prec+rec)>0 else 0)
    macro_f1 = sum(f1s)/len(f1s)
    inter = np.sum(pred_mask & gold_mask)
    union = np.sum(pred_mask | gold_mask)
    iou = inter/union if union>0 else 0
    dice = 2*inter/(np.sum(pred_mask)+np.sum(gold_mask)) if (np.sum(pred_mask)+np.sum(gold_mask))>0 else 0
    return {'accuracy': acc, 'f1_macro': macro_f1, 'iou': iou, 'dice': dice}

Common pitfalls

• Data splits are strictly performed at the slide/patient level to prevent information leakage from overlapping patches derived from the same WSI.
• WSIs used to derive patches for the classification dataset are explicitly excluded from the test set to avoid confounding the end-to-end pipeline evaluation.
• Inference uses overlapping sliding windows, requiring careful stitching to avoid boundary artifacts that could skew mask continuity and metric scores.

Evidence (verbatim from paper)

For the patch-level seven-class classification task, we evaluate model performance using Accuracy (Acc), Precision, Recall and F1-score. For whole-slide image (WSI) region segmentation, we report Intersection-over-Union (IoU) ,Dice coefficient (Dice), Precision, Recall and F1-score as the primary evaluation metrics.

Citation

@misc{xiao2026hepatobench,
  title={A Digital Pathology Resource for Liver Cancer Quantification with Datasets, Benchmarks, and Tools},
  author={Xiao et al. (2026)},
  year={2026},
  note={arXiv:2604.22858}
}

• arXiv: 2604.22858