Kaphathy/Dataset

MM-OphBench: Multi-Center Multimodal Clinical Ophthalmic Benchmark Dataset

<div align="center">

A Large-Scale, Standardized Multi-Center Benchmark Covering 7 Imaging Modalities & 4.3M+ Clinical Records

  ![Modalities]() ![Images]()

</div>

1. Executive Summary & Repository Overview

The MM-OphBench repository hosts a petabyte-scale, clinically harmonized ophthalmic image archive compiled from leading ophthalmic hospitals and benchmark cohorts. It spans 4,307,415 high-resolution diagnostic images and multimodal pairs across 7 distinct imaging modalities, with comprehensive clinical metadata, expert lesion segmentations, and hierarchical diagnostic taxonomies.

All archives are compressed into high-speed Zstandard chunks (.tar.zst) with pre-indexed SHA-256 manifests. A centralized, fully de-identified master index is provided in metadata/adapter_manifest_reviewed.parquet (39.9 MB), allowing instant cross-cohort querying across imaging modalities, device models, scan protocols, and clinical diagnoses without requiring massive raw downloads.

2. Multi-Center Clinical Provenance & Ethical Governance

All patient identifiable information (PII) including patient names, hospital record numbers (MRNs), national identity codes, and date-of-birth timestamps have been removed or replaced with deterministic SHA-256 cryptographic hashes (patient_hash, study_id), strictly complying with HIPAA Safe Harbor and GDPR pseudonymization principles.

Key Contributing Clinical Centers & Cohorts

1. 1.The Second Affiliated Hospital of Zhejiang University School of Medicine (浙二眼科)
2. 2.Contributions: Longitudinal FFA/ICGA dynamic angiography sequences (Final_eryuan_FFA), 17-class retinal disease benchmark (eryuanFundusT17RDCls), anterior segment photography.
3. 3.Clinical Leads: Certified ophthalmologists (Prof. Wang Guoping and retinal angiography specialist team).
4. 4.Beijing Tongren Hospital, Capital Medical University (北京同仁医院)
5. 5.Contributions: Longitudinal multi-condition clinical cohorts (tongren_archives), 66-class fine-grained disease screening benchmark (trhd3_FRD_Cls66_Train), and the standard 95-disease clinical taxonomy (tongren95_taxonomy.csv).
6. 6.The Second Xiangya Hospital of Central South University (中南大学湘雅二医院)
7. 7.Contributions: High-frequency ocular B-scan ultrasonography with detailed diagnostic text reports (BUltrasound).
8. 8.InEye Hospital of Chengdu University of TCM (成都中医药大学附属银海眼科医院)
9. 9.Contributions: Comprehensive Ultra-Widefield (UWF) Optos color imaging (yinhai_uwf, 274.6 GB), FFA dynamic sequences, and ocular B-scans.
10. 10.Chengdu First People's Hospital (成都市第一人民医院)
11. 11.Contributions: 50MHz Ultrasound Biomicroscopy (UBM) ciliary body and anterior chamber angle quantitative morphometry (ChengduShiyiUBM).
12. 12.Handan Eye Hospital (河北邯郸眼科医院)
13. 13.Contributions: Fluorescein angiography paired cohorts (handan_ffa, handan_ffa2) and epidemiological diabetic retinopathy screenings.
14. 14.Zhongshan Hospital, Fudan University (复旦大学附属中山医院)
15. 15.Contributions: Multimodal cohort comprising 419,014 clinical visits with cross-system validation.

3. Optical Modalities & Hardware Instrumentation

The dataset captures full optical, acoustic, and angiography representations across the anterior and posterior segments of the human eye:

4. Archive Directory & Shard Inventory

Archives are organized into self-contained sub-directories chunked at ~1024 MiB boundaries. Each shard is compressed using zstd -3 for fast decompressibility and high compression efficiency:

5. Master Metadata (metadata/) & Clinical Crosswalks

Instead of downloading multi-hundred-gigabyte archives to inspect sample attributes, download the lightweight parquet files in metadata/:

1. metadata/adapter_manifest_reviewed.parquet (39.9 MB)

Contains 4,307,415 rows mapping every asset with standard schema:

import duckdb

con = duckdb.connect()
# Query device distribution across modalities
con.execute("""
    SELECT modality_subtype, device_vendor, device_model, count(*) AS count
    FROM 'metadata/adapter_manifest_reviewed.parquet'
    GROUP BY 1, 2, 3
    ORDER BY count DESC
    LIMIT 10
""").df()

2. metadata/tongren95_taxonomy.csv (12 KB)

Comprehensive 95-class disease taxonomy mapping local clinical diagnoses (raw_zh) to standardized English terms (canonical_en), abbreviations (AMD, PDR, RVO, CSC), disease families (family_en), and specificity levels.

3. metadata/labeled_assets_summary.json (2.0 KB)

Complete breakdown of 1,817,159 instruction-tuned samples and 1,459,879 downstream high-confidence diagnostic splits.

6. Benchmark Evaluation Tasks

1. 1.17-Class Retinal Disease Classification (`eryuanFundusT17RDCls`): Standardized 17-class classification for common and sight-threatening retinal conditions (Normal, DR, AMD, RVO, Pathological Myopia, Macular Hole, Epiretinal Membrane, Retinal Detachment, Retinitis Pigmentosa, etc.).
2. 2.66-Class Multi-Center Fine-Grained Diagnosis (`trhd3_FRD_Cls66_Train`): Challenging 66-class classification covering rare and subtle ophthalmic pathologies validated across Beijing Tongren and Handan populations.
3. 3.95-Class Clinical Semantic Classification (`tongren16G95DFundus`, `tongren95_taxonomy`): Full-spectrum hospital diagnostic classification aligning free-text clinical reports and fundus photographs across 16 major disease families and 95 clinical subcategories.
4. 4.Dynamic FFA Vessel & Microvascular Leakage Segmentation (`Final_eryuan_FFA`, `handan_ffa`): Temporal sequence segmentation across early, arteriovenous, and late phases.
5. 5.Quantitative Anterior Chamber Morphometry & Angle-Closure Risk (UBM): Automated measurement of Anterior Chamber Depth (ACD in mm), Trabecular-Iris Angle (TIA), and angle-closure risk assessment across ChengduShiyiUBM, YinhaiUBM, and UBM_SW3200L.
6. 6.Ophthalmic Image Quality & Pre-Filtering Assessment (`FQ_Datasets`): Automated detection of non-diagnostic fundus images, lens opacity, uneven illumination, and motion artifacts for reliable clinical pipeline pre-filtering.
7. 7.Dual-View Diabetic Retinopathy Grading (`deepdrid`): Multi-field consistency and lesion grading combining macular and optic disc perspectives.
8. 8.Multi-Center Glaucoma Screening & Cup-to-Disc Ratio Estimation (`kaggle_glaucoma_v4`, `DGOCF`): Optic nerve head segmentation, vertical cup-to-disc ratio (vCDR) calculation, and multi-device glaucoma detection.

7. Quick Start: Extraction & Usage Guide

A. Download & Extract Archives (Linux / macOS)

To extract multi-part .tar.zst archives with maximum parallelism:

# Install zstd if not present
sudo apt-get install -y zstd  # Ubuntu/Debian
brew install zstd             # macOS

# Extract all shards for a specific dataset (e.g., BUltrasound)
cat archives/BUltrasound/BUltrasound-*.tar.zst | tar -I zstd -xvf - -C /path/to/destination/

B. Python Fast Metadata Query

import pandas as pd

# Read metadata directly from Hugging Face or local path
df = pd.read_parquet("metadata/adapter_manifest_reviewed.parquet", 
                     columns=["image_id", "modality_subtype", "device_model", "hospital_domain"])

# Filter for SS-OCT B-scans from Topcon Triton
topcon_oct = df[(df["modality_subtype"] == "oct_bscan") & (df["device_model"] == "DRI OCT Triton")]
print(f"Loaded {len(topcon_oct)} Topcon SS-OCT B-scans")

8. Citation & Acknowledgments

If you use this benchmark or any subsets in your research, please cite:

@dataset{mm_ophbench2026,
  title={MM-OphBench: A Multi-Center Multimodal Clinical Ophthalmic Benchmark Dataset},
  author={Zhejiang University, Beijing Tongren Hospital, InEye Hospital, Xiangya Hospital, and Contributors},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/Kaphathy/Dataset}}
}

9. License & Terms of Use

• —The aggregated metadata and clinical annotations are made available under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).
• —Research use only. Not for direct diagnostic or medical intervention without regulatory clearance.