README.md

metadata

license: cc-by-nc-sa-4.0
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - pathology
  - histopathology
  - benchmark
  - nucleus-counting
pretty_name: PathoVernier
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: pathovernier.jsonl
extra_gated_prompt: >-
  PathoVernier is released for non-commercial research use under CC BY-NC-SA
  4.0. The questions and reference counts are derived from Lizard, PUMA,
  PanNuke, CoNSeP and NuCLS; images are not redistributed and must be obtained
  from the original providers under their own terms. Access requests are
  reviewed manually.
extra_gated_fields:
  Full name: text
  Affiliation: text
  Country: country
  Intended use: text
  I will use PathoVernier for non-commercial research only and follow the terms of the source datasets: checkbox
  I will not use PathoVernier or models evaluated on it for clinical decision-making: checkbox

PathoVernier

PathoVernier is a benchmark for quantitative cell-composition reasoning on H&E histopathology patches. Each question requires counting specific nucleus types in specific image regions and deriving an answer from those counts with a deterministic rule. Reference counts come from expert nucleus annotations, so a model's reported counts can be checked in addition to its final answer.

• 759 questions on 553 256x256 patches, after review by four board-certified pathologists (27 of 786 candidates removed).
• Evaluation code: https://github.com/ChyaZhang/ASPECT
• Paper: https://arxiv.org/abs/2609.34277

Tasks

Task (skill)	Question
Region selection (argmax_region)	Which equal-area region has the highest density of a nucleus type?
Region comparison (region_compare)	How does one nucleus type compare between two equal-area regions (much fewer / comparable / much more)?
Cell-type comparison (type_compare)	How do two nucleus types compare over the whole patch?
Multi-step composition (multi_hop)	Select the region with the highest density of a type, then give that type's proportion band there.

Regions are quadrants, four horizontal bands or four vertical bands. Comparison categories: much fewer if r <= 0.5, much more if r >= 2.0, comparable otherwise (r = a/b; much more if b = 0 and a > 0). Proportion thresholds are stated in each question.

Source	Region selection	Region comparison	Cell-type comparison	Multi-step	Total
Lizard	82	85	135	98	400
PUMA	45	47	10	40	142
PanNuke	35	36	33	35	139
CoNSeP	17	7	13	7	44
NuCLS	11	11	2	10	34
Total	190	186	193	190	759

Files

File	Content
pathovernier.jsonl	one question per line (fields below)
patch_index.csv	source dataset, source image, crop origin (x0, y0), crop size, native image size, pixel size and nucleus count of each patch
reconstruct_images.py	rebuilds the 553 patches from the original datasets

Fields of pathovernier.jsonl: item_id, patch_id, dataset, organ, skill, partition, region, asked (queried nucleus type or type pair), question, options, answer, quantities (reference counts required by the question; multi-step questions also give _num / _den for the proportion), n_nuclei, extent_um (field of view).

Images

Images are not redistributed. Download the source datasets and run:

python reconstruct_images.py --lizard <Lizard> --pannuke <PanNuke> --consep <CoNSeP> \
    --puma <PUMA> --nucls <NuCLS> --out images/

Each argument is the extracted official release; source images are located by file name. The script writes images/<dataset>/<patch_id>.png (256x256 crops at native resolution), the layout used by the evaluation code.

Source	License	Used subset
Lizard	CC BY-NC-SA 4.0	DPath, CRAG and GlaS images
PUMA	CC0 1.0	nuclei and tissue ROIs
PanNuke	CC BY-NC-SA 4.0	folds 1-3
CoNSeP	provider terms, non-commercial research	train and test
NuCLS	provider terms	FOVs with at most 10% unlabeled nuclei

Metrics

• Acc: final-answer accuracy over all 759 questions.
• CA: agreement between the decision derived from the model's reported counts and the reference answer, over questions where that decision is determinable (multi-step: region only).
• RAWR: 1 - mean fraction of required counts within tolerance max(1, 0.1q), over correct answers that report all required counts.
• Count Acc: mean fraction of required counts within tolerance over all 759 questions.

The scorer is python -m aspect.score_pathovernier in the GitHub repository.

Citation

@article{zhang2026see,
  title   = {See, Measure, and Reason: Learning Visually Grounded Reasoning in Pathology},
  author  = {Zhang, Chengyang and Zhang, Wenchuan and Li, Bo and Li, Mengran and Liu, Xinyu and Yang, Jiaming and Chen, Jie and Zhang, Zhang and Yi, Yuhao and Bu, Hong and Lv, Jiancheng},
  journal = {arXiv preprint arXiv:2609.34277},
  year    = {2026}
}