Datasets:

Mikezcy
/

PathBind

license: cc-by-nc-4.0
task_categories:
  - visual-question-answering
  - object-detection
language:
  - en
  - zh
tags:
  - pathology
  - medical-imaging
  - benchmark
  - vqa
  - grounding
size_categories:
  - 1K<n<10K
pretty_name: PathBind
configs:
  - config_name: PathBind-VQA
    data_files:
      - split: manifest
        path: PathBind-VQA/manifest.tsv
  - config_name: PathBind-PTA
    data_files:
      - split: eval
        path: PathBind-PTA/pathbind_pta_600.tsv
  - config_name: PathBind-Grounding
    data_files:
      - split: manifest
        path: PathBind-Grounding/manifest.jsonl
extra_gated_prompt: >-
  PathBind is released for non-commercial research use only. It is NOT certified
  for clinical diagnostic use. By requesting access you agree to (1)
  non-commercial use, (2) no redistribution of the manifest, images, or
  derivatives, (3) citing the PathBind paper in any publication that uses the
  dataset, (4) obtaining the underlying source datasets (PathMMU, Path-VQA,
  Quilt-VQA, MedXpert-Path, OmniMed-Bright, PathVG) under their own licenses
  when using the VQA / Grounding manifests, and (5) not using PathBind-PTA to
  inform patient care. Fill in the fields below; we aim to review within five
  business days.
extra_gated_fields:
  Full name: text
  Affiliation: text
  Email (institutional preferred): text
  Country: country
  Intended use: text
  I agree to use PathBind for non-commercial research only: checkbox
  I agree not to redistribute the manifest, images, or derivatives: checkbox
  I understand PathBind-PTA is not for clinical use: checkbox
extra_gated_button_content: Request access

PathBind

A diagnostic benchmark for evaluating pathology vision-language models.

PathBind bundles 2,600 samples across three components — each filtered by an automated pipeline and finalized under expert pathologist review — to jointly probe visual dependence, cross-domain replication, and entity-level visual-semantic binding.

Config	# samples	Format
`PathBind-VQA`	1,500	sample-ID manifest (TSV)
`PathBind-PTA`	600	self-contained TSV, base64 JPEGs inline
`PathBind-Grounding`	500	sample-ID manifest (JSONL)

Diagnostic dimensions

Dim.	Name	Definition
D1	Coarse tissue/organ recognition	Broad tissue type, organ context, general histologic category.
D2	Cellular morphology	Nuclear atypia, cytoplasm, mitotic figures, necrosis.
D3	Cell–stroma interaction	Infiltration pattern, inflammatory distribution, desmoplasia.
D4	Spatial localization	Spatial position, arrangement, or distribution of structures.
D5	Diagnostic reasoning	Integrated visual evidence for diagnosis, grading, staging.
D6	Staining & IHC interpretation	Staining patterns, IHC positivity, intensity/distribution.

Repository layout

PathBind/
├── PathBind-VQA/           # 1,500 sample IDs (D1–D6 × 250)
│   └── manifest.tsv
├── PathBind-PTA/           # 600 self-contained samples (D1–D5 × 120)
│   └── pathbind_pta_600.tsv
└── PathBind-Grounding/     # 500 sample IDs (testA 380 + testB 120)
    └── manifest.jsonl

Usage

PathBind-VQA

We ship only sample IDs. Obtain the five source benchmarks yourself, then materialize the full TSV with the PathBind evaluation repo:

git clone https://github.com/<org>/PathBind
python pathbind/data/PathBind-VQA/build.py \
    --pathmmu_dir  /path/to/PathMMU  \
    --pathvqa_dir  /path/to/Path-VQA \
    --quiltvqa_dir /path/to/Quilt-VQA \
    --omnimed_dir  /path/to/OmniMed-Bright \
    --medxpert_dir /path/to/MedXpert-Path

Fields: index, dimension, source_dataset, source_sample_id, answer_type.

Composition: PathMMU 1,311 · OmniMed-Bright 102 · Path-VQA 37 · Quilt-VQA 25 · MedXpert-Path 25.

PathBind-Grounding

Similarly, obtain PathVG and materialize:

python pathbind/data/PathBind-Grounding/build.py --pathvg_dir /path/to/PathVG

Fields: split, image_id, bbox_id, expression, bbox, width, height.

PathBind-PTA

Self-contained — no external download needed. Follows the PathoSage TSV convention: the image column contains base64-encoded JPEG data (long side ≤ 1024 px, quality 85).

Balanced across five diagnostic dimensions (D1–D5, 120 each) and answer options (A–D, 150 each).

Running an evaluation

git clone https://github.com/<org>/PathBind
cd PathBind
python -m pathbind.eval_a1 \
    --data PathBind-VQA/pathbind_vqa_1500.tsv \
    --model_module pathbind.models.patho_r1 --ckpt /path/to/Patho-R1 \
    --out results/vqa.jsonl

Replace --data with the PTA or Grounding materialized file for the other two components. See the PathBind code repo README for the full API and how to port a new model.

Citation

Coming soon.

License

PathBind-VQA and PathBind-Grounding manifests: MIT. Referenced source datasets remain under their original licenses — please follow the terms of PathMMU / Path-VQA / Quilt-VQA / MedXpert-Path / OmniMed-Bright / PathVG when using them.
PathBind-PTA component: CC BY-NC 4.0 (non-commercial research use only, not for clinical use).