RadM-Bench: A Bilingual Multimodal Benchmark for Diagnostic Radiology

📖 Overview

RadM-Bench is a bilingual (English–Chinese) multimodal benchmark for evaluating the diagnostic performance of multimodal large language models (MLLMs) in radiology. It is built to expose three blind spots in existing benchmarks: the gap between curated 2D snapshots and real volumetric (3D) imaging, the gap between public teaching cases and routine clinical practice, and the gap between English-only and bilingual clinical contexts.

The benchmark comprises 720 cases, evenly distributed across nine radiological subspecialties (80 cases each), and split into two complementary subsets of 360 cases. Each case pairs a clinical history with radiologist-curated key images, and — for the clinical subset — full volumetric cine clips, enabling models to be evaluated under progressively richer inputs.

🏥 Dataset Composition

Subsets

• RadEdu (N=360) — English public teaching cases curated from Eurorad (Aug 16, 1999 – May 15, 2024). From 9,640 eligible studies, a stratified random sample of 360 (40 per subspecialty) was drawn. Rare-disease status follows the NORD catalogue. Rare-disease–enriched (45.6% rare).
• RealClin (N=360) — Chinese routine clinical cases from a Chinese academic tertiary hospital (Jan 1, 2022 – Jan 1, 2024), retrospectively and randomly sampled from daily practice with the same modality and subspecialty-balanced criteria. Rare-disease status follows the Chinese National Health Commission Rare Disease Directory. Reflects routine prevalence (8.6% rare). All data de-identified via DICOM tag removal and manual redaction.

Subspecialties (80 cases each)

Abdominal · Breast · Cardiovascular · Chest · Head and Neck · Musculoskeletal · Neuroradiology · Pediatric · Urogenital

Imaging Modalities

Modality	Overall	RadEdu	RealClin
CT	250 (34.7%)	81 (22.5%)	169 (46.9%)
Hybrid (multi-modality)	299 (41.5%)	207 (57.5%)	92 (25.6%)
MR	163 (22.6%)	67 (18.6%)	96 (26.7%)
DR (Radiography)	8 (1.1%)	5 (1.4%)	3 (0.8%)

Disease Rarity

	Overall	RadEdu	RealClin
Common (0)	525 (72.9%)	196 (54.4%)	329 (91.4%)
Rare (1)	195 (27.1%)	164 (45.6%)	31 (8.6%)

The contrast is intentional: RadEdu over-represents rare and pedagogically valuable pathologies, while RealClin mirrors the common-disease spectrum of routine clinical work — allowing the "rarity premium" of teaching benchmarks to be measured directly.

🎛️ Multimodal Inputs

Each case can be presented to a model under progressively enriched conditions:

1. Clinical history only — text presented in the case's native language (English for RadEdu, Chinese for RealClin).
2. History + 2D key images — radiologist-selected slices exported as PNG/JPEG at clinically appropriate window/level (all 720 cases).
3. History + volumetric data — for the 360 RealClin cases, complete CT/MR series exported as canonical MP4 cine clips (512×512, H.264 High Profile, 10 fps). Multi-sequence MR and multiphase CT are stored as separate clips per sequence/phase.

🗂️ Data Format

Each case in data.json is a JSON object:

{
    "ID": "RealClin_001",
    "DataSource": "RadEdu | RealClin",
    "Subspecialty": "medical_subspecialty",
    "Modality": "CT | MR | Hybrid | DR",
    "DiseaseRarity": 0,
    "ClinicalHistory": "detailed clinical history text",
    "Answer": "reference_diagnosis_1;acceptable_alternative_2",
    "num_image": 1,
    "image_paths": ["./images/RealClin/RealClin_001-1.jpg"],
    "video_paths": ["./videos/RealClin/RealClin_001/MOVIE-0010.mp4"]
}

Field Descriptions

• ID — unique case identifier (RadEdu_001–RadEdu_360, RealClin_001–RealClin_360)
• DataSource — subset: RadEdu or RealClin
• Subspecialty — radiological subspecialty
• Modality — primary imaging modality (CT, MR, Hybrid, DR)
• DiseaseRarity — 0 = common disease, 1 = rare disease
• ClinicalHistory — patient presentation and relevant history (native language)
• Answer — reference diagnosis(es), corresponding to the paper's comprehensive diagnosis; diagnoses are separated by semicolons (;), and any listed diagnosis is accepted during evaluation
• num_image — number of 2D key images
• image_paths — list of key-image file paths
• video_paths — paths to volumetric data (RealClin only): the presence of this field indicates a case carries volumetric (3D) measurement data. Most entries are MP4 cine clips; for a small number of DR (radiography) cases where no volumetric series exists, a key image (.jpg) is provided here instead.

⬇️ How to Use

The full dataset (metadata, all key images, and volumetric clips) is hosted in this repository. Download the whole repo:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Elizabeth123/RadM-Bench",
    repo_type="dataset",
    local_dir="./RadM-Bench"
)

Then load the cases:

import json

with open('./RadM-Bench/data.json', 'r', encoding='utf-8') as f:
    data = json.load(f)

for item in data:
    print(item['ID'], item['DataSource'], item['Modality'])
    print(item['ClinicalHistory'])
    print(item['Answer'])
    print(item['image_paths'])
    print(item.get('video_paths', []))

The relative image_paths / video_paths in data.json resolve directly against the downloaded folder.

Mirror: the dataset is also available on Google Drive.

🗂️ Repository Structure

RadM-Bench/
├── data.json                 # 720 cases (metadata + paths)
├── README.md
├── images/                   # key images
│   ├── RadEdu/
│   └── RealClin/
├── videos/                   # volumetric clips
│   └── RealClin/<ID>/
└── article_images/           # figures from the paper

🔒 Ethics

This study was approved by the Institutional Review Board of Henan Provincial People's Hospital (approval no. 2025163) and conducted in line with the Declaration of Helsinki. It used retrospective de-identified clinical data and publicly available educational radiology cases for secondary analysis; informed consent was waived owing to the retrospective design and full anonymization. All images in the repository are anonymized with no identifiable individual features.

Citation

If you find this dataset useful in your research, please cite our paper:

@article{radmbench2025,
  title={RadM-Bench: A Bilingual Benchmark for Evaluating Diagnostic Performance of Multimodal Large Language Models in Radiology},
  author={Wu, Qingxia and Wu, Qingxia and Zhang, Peipei and others},
  journal={JMIR},
  year={2026}
}