MedRCube

📖 arXiv Paper • 💻 GitHub

Overview

MedRCube is a multidimensional medical imaging benchmark designed to answer not just how well a model performs, but where, why, and how credibly it does so.

It comprises 7,626 rigorously constructed samples from 36 datasets, spanning 5 anatomical regions (Heart, Chest, Breast, Lung, Brain), 4 imaging modalities (X-ray, CT, MRI, Ultrasound), and 8 cognitive tasks organized into a three-tier hierarchy, built through a systematic pipeline with radiologist and clinical expert participation throughout.

Every sample is mapped into a structured Competency Space defined by three orthogonal axes (Anatomy × Modality × Task). Each intersection forms a Competency Voxel for precise capability localization. By constructing multi-level task chains on the same image, MedRCube further enables reasoning credibility verification — checking whether a model's correct diagnosis is genuinely supported by correct perception, or merely a lucky guess.

Data Fields

Each record in test.json contains:

Field	Description
id, dataset	Sample identifier and source dataset name
image_path	Relative path to the image (if present)
question	The question text
option_A / option_B / option_C / option_D	answer choices
gt_answer, correct_index	Ground-truth answer and its index
task, modality, parts	Competency Space coordinates
original_task	Original task in source dataset
restricted	true if images cannot be redistributed directly

Loading with datasets

from datasets import load_dataset

ds = load_dataset("Flmc/MedRCube", split="test")
print(ds[0])

If an example is restricted (or the image file does not exist), its image field will be None.

Download for Evaluation

To use with the GitHub evaluation scripts, download the full snapshot:

huggingface-cli download YOUR_HF_ORG/MedRCube \
  --repo-type dataset \
  --local-dir ./MedRCube \
  --local-dir-use-symlinks False

Then point the evaluator's --dataset_path to the snapshot root.

Restricted Sources

Some sources cannot redistribute images. We release the questions now, and will provide reproducible preprocessing scripts (coming soon) so researchers can reconstruct images after obtaining access from the original providers. See restricted_sources.json for the full list.

License

This dataset is released under CC-BY-NC 4.0. The evaluation code is licensed under Apache-2.0.

Citation

If you find MedRCube helpful, please cite:

@misc{medrcube2026,
  title={MedRCube: A Multidimensional Framework for Fine-Grained and In-Depth Evaluation of MLLMs in Medical Imaging},
  author={Bao, Zhijie and Chen, Fangke and Bao, Licheng and Zhang, Chenhui and Chen, Wei and Peng, Jiajie and Wei, Zhongyu},
  journal={arXiv preprint},
  year={2026},
  eprint={2604.13756},
  url={https://arxiv.org/abs/2604.13756},
}