praiselab-picuslab/MMMED

🏥 Multilingual Multimodal Medical Exam Dataset for Visual Question Answering in Healthcare

The Multilingual Multimodal Medical Exam Dataset (MMMED) is a comprehensive benchmark designed to evaluate Vision-Language Models (VLMs) on medical multiple-choice question answering (MCQA) tasks. This dataset combines medical images and multiple-choice questions in Spanish, English, and Italian, derived from the Médico Interno Residente (MIR) residency exams in Spain.

The dataset includes challenging, real-world medical content, with images from various diagnostic scenarios, making it ideal for assessing VLMs in cross-lingual medical tasks.

🔒 How to Access the Dataset

You can access the MMMED dataset via Hugging Face. Follow these steps to download it:

from datasets import load_dataset

# Login using e.g. `huggingface-cli login` to access this dataset
ds = load_dataset("praiselab-picuslab/MMMED")

🌟 Key Features:

• Languages: 🇪🇸 Spanish, 🇬🇧 English, 🇮🇹 Italian
• Medical Content: Questions based on real Spanish residency exams
• Image Types: Diagnostic medical images (e.g., CT scans, X-rays)
• Categories: 24 medical specialties (e.g., Digestive Surgery, Cardiology)
• Multimodal: Each question comes with a medical image 📸

🛠️ Dataset Workflow

Here is the general workflow for building the MMMED dataset for Vision-Language Model (VLM) evaluation:

📊 Dataset Overview

The MMMED dataset contains 194 questions from the MIR exams and features images from real-world medical contexts. The dataset is organized into 24 medical categories, each with corresponding textual questions and image-based choices.

Statistic	🇪🇸 Spanish	🇬🇧 English	🇮🇹 Italian
# Questions	194	194	194
# Categories	24	24	24
Last Update	2024	2024	2024
Avg. Option Length	6.85	6.57	6.71
Max. Option Length	41	39	39
Total Question Tokens	10,898	10,213	10,545
Total Option Tokens	5,644	5,417	5,528
Avg. Question Length	56.18	52.64	54.36
Max. Question Length	223	190	197

🖼️ Image Types

Categorization of Image Types in the MMMED Dataset. This figure presents the four main categories of images included in the dataset and their respective distributions.

✨ Example MMCQA

Each multimodal multiple-choice question-answer (MMCQA) pair integrates three essential components with the following structure:

• Category: C
• Question: Q
• Image URL: I
• Answer Options: O
• Correct Answer: 💡

Here’s an illustrative example of multimodal QA in three languages:

🔍 List of Open-Source and Closed-Source Vision-Language Models (VLMs) Used

This table shows the parameter sizes, language models, vision models, and average scores of VLMs evaluated on the OpenVLM Leaderboard.

Rank	Method	Param (B)	Language Model	Vision Model	Avg Score (%)
Open-Source Models
167	PaliGemma-3B-mix-448	3	Gemma-2B	SigLIP-400M	46.5
108	DeepSeek-VL2-Tiny	3.4	DeepSeekMoE-3B	SigLIP-400M	58.1
135	Phi-3.5-Vision	4	Phi-3.5	CLIP ViT-L/14	53.0
209	LLaVA-v1.5-7B	7.2	Vicuna-v1.5-7B	CLIP ViT-L/14	36.9
Closed-Source Models
34	Claude3.5-Sonnet-20241022	Unknown	Closed-Source	Closed-Source	70.6
24	GPT-4o (1120, detail-high)	Unknown	Closed-Source	Closed-Source	72.0
20	Gemini-2.0-Flash	Unknown	Closed-Source	Closed-Source	72.6

📈 VLM Performance on MMMED

The following figure presents the accuracy of different VLMs in each language tested:

🌐 Notes

Dataset Usage: The dataset is intended for academic and research purposes only. It is not recommended for clinical decision-making or commercial use.

👨‍💻 This project was developed by Antonio Romano, Giuseppe Riccio, Mariano Barone, Gian Marco Orlando, Diego Russo, Marco Postiglione, and Vincenzo Moscato
University of Naples, Federico II

📝 License

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.