Histopath C Eval

histopath-c-eval

Histopath-C: Towards Realistic Domain Shifts for Histopathology Vision-Language Adaptation — Mehrdad Noori et al. (2026) (arXiv:2601.12493, 2026)

What this evaluates

Evaluates the robustness of vision-language models (VLMs) and test-time adaptation (TTA) methods when applied to histopathology images under realistic domain shifts. It probes how well models maintain classification accuracy when exposed to synthetic corruptions like staining variations, dust, blurring, and noise that mimic real-world clinical imaging artifacts.

Datasets

• NCT-7K — total 7000; splits: test (-1)
• NCT-100K — total 100000; splits: test (-1)
• LC25000 — total 25000; splits: Lung (-1), Colon (-1), All (-1)
• SkinCancer — total ?; splits: test (-1)
• RenalCell — total ?; splits: test (-1)
• MHIST — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly predicted class labels out of the total number of test instances. Calculated as (number of correct predictions / total predictions) * 100.

Input / output format

Input: Histopathology image patches or tiles (typically 224x224 or 448x448 pixels depending on the VLM backbone) paired with a ground-truth class label. During evaluation, images are subjected to one of ten synthetic corruptions (e.g., stain-light, dust, defocus blur, Gaussian noise).

Output: Predicted class label (or probability distribution over classes) for each input image.

Scoring recipe

def calculate_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100

Common pitfalls

• Entropy-based TTA methods (e.g., TENT, TPT) frequently degrade performance on histopathology data compared to natural images, contrary to typical benchmarks.
• CLIPArTT is explicitly noted as inapplicable to datasets with fewer than three classes.
• All reported results are averaged over three independent runs, with standard deviations provided in supplementary material.
• The benchmark evaluates both clean and corrupted ('Dataset-C') versions of each dataset, requiring separate evaluation passes.

Evidence (verbatim from paper)

To assess the effectiveness and generalization capabilities of LATTE, we perform extensive evaluations across a suite of diverse and challenging histopathology datasets... For each, we apply the ten corruptions in Histopath-C, denoted as Dataset-C. ... Specifically, using four templates improves accuracy, whether through text averaging or loss averaging.

Citation

@misc{noori2026histopathc,
  title={Histopath-C: Towards Realistic Domain Shifts for Histopathology Vision-Language Adaptation},
  author={Mehrdad Noori et al. (2026)},
  year={2026},
  note={arXiv:2601.12493}
}

• arXiv: 2601.12493