Euanyu/LMOD-Plus-Sample

LMOD+ — Sample Subset

A 1,076-instance sample of LMOD+, a large-scale multimodal ophthalmology benchmark for developing and evaluating multimodal large language models (MLLMs).

This repository is a preview subset intended for quickly inspecting the data format, prototyping evaluation harnesses, and running smoke tests. The full benchmark contains 32,633 instances across 12 ophthalmic conditions and 5 imaging modalities.

📄 Paper: ACM Transactions on Computing for Healthcare (2026) · arXiv:2509.25620 🌐 Project page: https://kfzyqin.github.io/lmod_plus/

Quick start

from datasets import load_dataset

# Anatomical recognition with bounding boxes
ds = load_dataset("Euanyu/LMOD-Plus-Sample", "origa", split="test")
print(ds[0]["bboxes"])
# [{'region_type': 'Optic Disc', 'xmin': 0.528, 'ymin': 0.400, 'xmax': 0.651, 'ymax': 0.626}, ...]

# Multi-class disease diagnosis
ds = load_dataset("Euanyu/LMOD-Plus-Sample", "eye_disease", split="test")
print(ds[0]["diagnosis"])  # 'cataract'

# Disease staging + demographics
ds = load_dataset("Euanyu/LMOD-Plus-Sample", "oimhs", split="test")
print(ds[0]["mh_stage"], ds[0]["patient_age"], ds[0]["patient_sex"])

Streaming works too, which is useful for the larger origa config:

ds = load_dataset("Euanyu/LMOD-Plus-Sample", "origa", split="test", streaming=True)

Configurations

Config	Instances	Modality	Primary task	Annotations
cataract_1k	200	Surgical scenes (SS)	Anatomical recognition	688 boxes over 15 region types
cau001	200	Lens photographs (LP)	Anatomical recognition	655 boxes (eye / iris / pupil)
oimhs	76	OCT	Staging + demographics	508 boxes, 206 masks, MH stage 1–4
origa	200	Color fundus (CFP)	Binary screening	400 boxes (disc / cup), CDR
eye_disease	400	Color fundus (CFP)	Multi-class diagnosis	4-class label
Total	1,076

Schema

All configs share a common core:

Field	Type	Description
image_id	string	Identifier, unique within a config
dataset	string	Source dataset name
modality	string	SS, LP, OCT, or CFP
image_type	string	Free-text modality description from the annotation pipeline
image	Image	The clean, un-annotated image
relative_path	string	Path to this image inside the raw archives (see below)

Config-specific fields:

• cataract_1k, cau001, origa, oimhs — bboxes: list of {region_type, xmin, ymin, xmax, ymax}
• origa — glaucoma_label (Glaucoma / Non-Glaucoma), cdr (cup-to-disc ratio), ecc_cup, ecc_disc
• oimhs — segmentations: list of {segmentation_class, mask} (retina / choroid); mh_stage (1–4); patient_id, patient_age, patient_sex, eye_side; image-quality flags low_signal_strength, signal_shield, image_blur
• eye_disease — diagnosis: cataract, diabetic_retinopathy, glaucoma, or normal

Bounding box convention

Coordinates are normalized to [0, 1] with the origin at the top-left corner. Convert to pixels by multiplying by the image width and height:

from PIL import ImageDraw

row = ds[0]
img = row["image"].convert("RGB")
W, H = img.size
draw = ImageDraw.Draw(img)
for b in row["bboxes"]:
    draw.rectangle([b["xmin"] * W, b["ymin"] * H, b["xmax"] * W, b["ymax"] * H], outline="lime", width=4)

eye_disease is a pure classification config and has no bboxes column.

Raw archives

The parquet files contain the clean images plus structured annotations, which covers most use cases. If you need the original on-disk layout — including the pre-rendered annotation overlays used to produce the paper's figures — each sub-dataset is also shipped as a tarball under raw/:

raw/subset_Cataract-1K.tar.gz    600 files    312 MB
raw/subset_CAU001.tar.gz         600 files    202 MB
raw/subset_OIMHS.tar.gz          716 files     78 MB
raw/subset_ORIGA.tar.gz          800 files   1403 MB
raw/subset_eye_disease.tar.gz    400 files     73 MB

from huggingface_hub import hf_hub_download
import tarfile

p = hf_hub_download("Euanyu/LMOD-Plus-Sample", "raw/subset_OIMHS.tar.gz", repo_type="dataset")
tarfile.open(p).extractall("./lmod_raw")

Each sample folder holds the clean image, an information.json (or KEYINFORMATION*.json), and an annotated/ or metadata/ subtree with overlays. The image_path field inside every JSON is relative to the subset root, so subset_OIMHS/ + 100_30/visualization.png resolves directly.

Tasks

LMOD+ supports four task families. This subset exercises all of them:

1. Anatomical recognition — localize structures (cornea, pupil, iris, optic disc/cup, retina, choroid, surgical instruments) via bounding boxes and segmentation masks.
2. Disease screening — binary detection, e.g. glaucoma vs. non-glaucoma in origa.
3. Disease staging — severity grading, e.g. macular hole stages 1–4 in oimhs.
4. Demographic prediction — infer patient age and sex from the image (oimhs).

Reported zero-shot performance in the paper is far from saturated: the strongest of the 24 evaluated MLLMs reaches roughly 58% accuracy on disease screening.

Images are unmodified

Image bytes are embedded bit-exact from the source pipeline — no resizing, re-encoding, or recompression. Note that this means origa inherits the pipeline's PNG conversion of originally-JPEG fundus photographs, which is why that config is disproportionately large (~700 MB for 200 images at 2518×2048).

Source datasets and attribution

LMOD+ is an aggregation of publicly released ophthalmic datasets, re-annotated with a unified multi-granular schema. Under CC BY 4.0 you must credit LMOD+; you are also responsible for honouring the terms of each upstream source, which are not superseded by this repository's license.

Sub-dataset	Modality	Upstream source
Cataract-1K	Surgical scenes	Cataract-1K cataract surgery dataset
CAU001	Lens photographs	See the LMOD+ paper, §Datasets
OIMHS	OCT	OIMHS — OCT images with macular hole segmentation
ORIGA	Color fundus	ORIGA-light retinal fundus dataset
Eye Diseases	Color fundus	4-class fundus classification collection; see the LMOD+ paper

If you redistribute or build on this subset, cite both LMOD+ and the relevant upstream dataset.

Demographics and privacy

The oimhs config includes patient_id, patient_age, patient_sex, and eye_side. These are carried over from the public upstream release and are required for the demographic-prediction task; patient_id is a pseudonymous integer from the source dataset and enables patient-level grouping so that images from the same patient do not straddle a train/test boundary. No direct identifiers are present. All internal file paths from the annotation pipeline have been stripped from both the parquet files and the raw archives.

Known issues

• Region-label typos. Two boxes in cataract_1k carry malformed labels: case5013_14 has a pupil1 box alongside its Pupil box, and case5014_18 has cornea1 alongside Cornea. These are preserved as-is so the data matches what was evaluated in the paper. Filter or normalize them if your pipeline is label-sensitive.
• Partial anatomical coverage in cau001. 50 of 200 samples annotate only the iris and pupil, without an enclosing Left Eye / Right Eye box.
• Class balance is by construction. The per-class counts in this subset (100 per diagnosis, 19 per macular hole stage) were sampled to be balanced and do not reflect prevalence in the full benchmark or in a clinical population.

Relationship to the full release

This subset is drawn from the full LMOD+ benchmark (32,633 instances; additionally covering Harvard FairSeg, CatDet2, REFUGE, IDRiD, and G1020). It is intended for format inspection and prototyping — do not report benchmark numbers on this subset, as the sample sizes are far too small for meaningful comparison and the class balance is artificial.

Citation

@article{10.1145/3801746,
  author     = {Qin, Zhenyue and Liu, Yang and Yin, Yu and Ding, Jinyu and Zhang, Haoran
                and Li, Anran and Campbell, Dylan and Wu, Xuansheng and Zou, Ke
                and Keenan, Tiarnal D. L. and Chew, Emily Y. and Lu, Zhiyong
                and Tham, Yih Chung and Liu, Ninghao and Zhang, Xiuzhen and Chen, Qingyu},
  title      = {LMOD+: A Comprehensive Multimodal Dataset and Benchmark for Developing
                and Evaluating Multimodal Large Language Models in Ophthalmology},
  year       = {2026},
  issue_date = {July 2026},
  publisher  = {Association for Computing Machinery},
  address    = {New York, NY, USA},
  volume     = {7},
  number     = {3},
  url        = {https://doi.org/10.1145/3801746},
  doi        = {10.1145/3801746},
  journal    = {ACM Trans. Comput. Healthcare},
  month      = jun,
  articleno  = {52},
  numpages   = {38}
}

The preliminary version of this benchmark appeared at NAACL 2025:

@inproceedings{qin-etal-2025-lmod,
  title     = {{LMOD}: A Large Multimodal Ophthalmology Dataset and Benchmark for Large Vision-Language Models},
  author    = {Qin, Zhenyue and Yin, Yu and Campbell, Dylan and Wu, Xuansheng and Zou, Ke
               and Liu, Ninghao and Tham, Yih Chung and Zhang, Xiuzhen and Chen, Qingyu},
  booktitle = {Findings of the Association for Computational Linguistics: NAACL 2025},
  month     = apr,
  year      = {2025},
  address   = {Albuquerque, New Mexico},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2025.findings-naacl.135/},
  doi       = {10.18653/v1/2025.findings-naacl.135},
  pages     = {2501--2522}
}