abhijit2k01/proton-benchmarks

PROTON / PROTON+ medical OOD benchmarks

Curated test streams for test-time out-of-distribution detection with medical vision-language models, released with PROTON+ (https://github.com/aj-das-research/PROTON, project page https://aj-das-research.github.io/PROTON/), the follow-up to PROTON (MICCAI 2026, https://arxiv.org/abs/2606.20913).

File	Content
ophthalmology.parquet, radiology.parquet, pathology.parquet	one row per image: set, label (ID class index or -1), class_name, source, path (as in the repo manifests), image (PNG bytes, long side <= 1024 px)
<benchmark>_features_<vlm>.parquet	the l2-normalised image embedding of every row from the frozen VLM used in the paper (FLAIR, UniMed-CLIP, QuiltNet)
<benchmark>_text_<vlm>.json	the class-prompt embeddings of every prompt set, the logit scale and the default prompt set
<benchmark>_manifest.csv	the seeded manifest without images

set is id (ID test images), id_bank (a disjoint labelled ID split, used only by the privileged baselines and for prompt selection), semantic, covariate or far. The radiology file also holds semantic_covid (COVID-19 radiographs), which is kept for completeness and not used in the paper.

Preview

The Hub viewer cannot display the image bytes, so here are two images per split of each benchmark, drawn with a fixed seed (frame colour: blue ID, orange semantic, red covariate, grey far).

Two example rows of ophthalmology.parquet, an ID image and a covariate-shifted one:

set	label	class_name	source	path	image
id	0	normal	fives	data/raw/fives/.../test/Original/151_N.png	PNG bytes, 1024 x 1024 px
covariate	-1	uwf	deepdrid	data/raw/deepdrid/.../ultra-widefield_images/.../100_l4.jpg	PNG bytes, 1024 x 807 px

Load one image (the pathology file is the smallest, 115 MB):

import io, pandas as pd
from huggingface_hub import hf_hub_download
from PIL import Image
df = pd.read_parquet(hf_hub_download("abhijit2k01/proton-benchmarks", "pathology.parquet", repo_type="dataset"))
row = df[df.set == "covariate"].iloc[0]
print(row.set, row.class_name, row.source); Image.open(io.BytesIO(row.image)).save("example.png")

Recompute the results

With the features and text embeddings every number of the paper can be recomputed on a CPU:

import pandas as pd, numpy as np, json
from proton import score_stream, ood_metrics
from proton.online import proton_plus_config
fe = pd.read_parquet("ophthalmology_features_flair.parquet"); tx = json.load(open("ophthalmology_text_flair.json"))
T = np.array([np.mean(tx["text"][tx["default_prompts"]]["embeddings"][c][:k], 0) for c, k in enumerate(tx["text"][tx["default_prompts"]]["counts"])])
stream = fe[fe.set.isin(["id", "covariate"])].sample(frac=1, random_state=0)
out = score_stream(np.stack(stream.embedding), T, tx["logit_scale"], config=proton_plus_config())
print(ood_metrics(out["score"], (stream.set != "id").astype(int)))

Citation

If you use these benchmarks, please cite PROTON and PROTON+, and the original datasets listed below:

@inproceedings{das2026proton,
  title={PROTON: Prototype-Based Test-Time Online OOD Detection for Medical VLMs},
  author={Das, Abhijit and Wasalathilaka, Nichula and Lu, Yifan and Dukre, Adinath and Mahapatra, Dwarikanath and Khan, Shadab and Razzak, Imran},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={527--537},
  year={2026},
  organization={Springer}
}
@misc{das2026protonplus,
  title  = {PROTON+: Contamination-Robust Prototype Banks for Test-Time Online OOD Detection with Medical VLMs},
  author = {Das, Abhijit and Wasalathilaka, Nichula and Lu, Yifan and Dukre, Adinath and Mahapatra, Dwarikanath and Khan, Shadab and Razzak, Imran},
  year   = {2026},
  note   = {Preprint, https://github.com/aj-das-research/PROTON}
}

Sources and licenses (please cite the original datasets)

• FIVES (Jin et al., Sci. Data 2022), CC BY 4.0. DeepDRiD ultra-widefield (Liu et al., Patterns 2022), CC BY-SA 4.0.
• Imagenette (fast.ai; a subset of ImageNet, Deng et al. 2009), for research use under the ImageNet terms.
• RSNA Pneumonia Detection Challenge (Shih et al., Radiology: AI 2019; images from NIH ChestX-ray14), non-commercial research use. NIH ChestX-ray14 sample (Wang et al., CVPR 2017), NIH Clinical Center release. Paediatric chest X-rays (Kermany et al., Cell 2018), CC BY 4.0.
• PatchCamelyon (Veeling et al., MICCAI 2018), CC0. Camelyon17-WILDS (Koh et al., ICML 2021), CC0. NCT-CRC-HE (Kather et al., 2018), CC BY 4.0.

The streams, splits and de-duplication rules are described in the PROTON+ paper and implemented in proton/data/manifests.py of the repository. Images are redistributed only to make the exact test streams reproducible; the original licenses apply.