Dataset Card for MedReaMM

MedReaMM is a multimodal medical diagnostic reasoning benchmark designed to evaluate the diagnostic capabilities of Large Language Models (LLMs) and Large Multimodal Models (LMMs) in real-world clinical scenarios. The benchmark dataset is collected from top-tier medical journals and curated clinical case report databases, and includes 625 expert-validated real-world cases.

Unlike existing benchmarks that primarily rely on textual modality, MedReaMM includes a large number of medical images (such as CT, X-ray, MRI, etc.), with an average of 2.79 images per case, totaling 1746 images. Each case in the dataset is provided with detailed patient information, images and their descriptions (captions), and standardized diagnoses annotated with ICD-11 codes.

This dataset aims to bridge the gap between text-based benchmarks and real-world clinical practice, providing a more comprehensive platform for evaluating and developing more powerful medical AI models.

Features

• Authenticity: All cases are sourced from real-world clinical reports and validated by experts, ensuring the data's reliability and clinical relevance.

• Multimodality: Each case includes textual information and various medical images, allowing models to process and understand data from different modalities simultaneously, which is closer to the actual diagnostic process.

• Standardized Diagnoses: The diagnostic results for each case are standardized and annotated with ICD-11 codes, facilitating objective evaluation and cross-model comparison.

• Complexity: The dataset not only includes common diseases but also covers a variety of rare diseases, which places higher demands on the model's medical knowledge and reasoning capabilities.

• Traceability: Each case provides source information, making it convenient for researchers to trace and verify.

Dataset File Structure

The dataset consists of a single .jsonl file and a folder named images.

benchmark.jsonl: This is a JSON Lines file where each line represents a single case, containing all the textual information for that case and referencing the corresponding image files via the path field. The file structure is as follows:

{"pmid": "...", "patient_info": {"basic_info": "...", "supplementary_info": [{"type": "fig", "id": 1, "subfig": null, "path": ["images/.../1.jpg"], "caption": "...", "detailed_caption": "...", "modalities": ["..."]}, ...]}, "diagnosis": "...", "standardized_diagnosis": [...], "year": 1976, "classification": "..."}

images/: This folder contains all the medical image files referenced by the cases. The image files are organized by pmid (the unique identifier for each case), for example:

images/
├── case-13544/
│   ├── fig/
│   │   ├── 1.jpg
│   │   ├── 2.jpg
│   │   └── ...
│   └── ...
├── 12345/
│   └── ...
└── ...

In the .jsonl file, the value of the path field is a relative path pointing to the corresponding image file within the images folder. Models need to load the images based on these paths for multimodal reasoning.