Datasets:
🧠 MM-NeuroOnco Image Collection
This repository hosts the MRI slice image collection and metadata used in the MM-NeuroOnco project.
MM-NeuroOnco is a large-scale multimodal benchmark and instruction dataset designed for clinically grounded MRI-based brain tumor diagnosis and interpretable reasoning.
📖 Overview
Unlike traditional classification datasets, MM-NeuroOnco emphasizes clinically interpretable reasoning across MRI physics, anatomical localization, and tumor morphology.
Figure 1: Overview of MM-NeuroOnco tasks and capabilities.
📊 Dataset Statistics
The dataset aggregates data from 20 publicly available sources and covers 8 tumor types.
Figure 3: Distribution of tumor types and question diversity in the instruction dataset.
🔍 Qualitative Examples
The dataset includes rich Chain-of-Thought (CoT) explanations and grounded diagnosis, supporting both open-ended and closed-ended VQA tasks.
Figure 7: Sample cases showing model reasoning, attribute extraction, and diagnosis.
⚙️ Pipeline
Our fully reproducible multi-stage label construction pipeline ensures high-quality silver labels through a "Skeptic-Conservative" dual-model approach.
Figure 2: The three-stage pipeline: Extraction, Fusion, and Audit.
⚠️ Important: BraTS Data Restriction
Note: Due to licensing restrictions, BraTS 2021 imaging data are NOT redistributed in this repository.
Users must obtain the BraTS dataset independently:
- Register at the official BraTS 2021 website.
- Agree to the data usage terms.
- Download the data.
Directory Alignment:
Once downloaded, the relevant BraTS images (typically in folders like 16_t1_t2_t1ce_flair_mask) must be placed into the appropriate directory structure to match the image_path fields in our JSON annotations.
📂 Directory Structure
The directory structure in this repository mirrors the image_path entries defined in the official benchmark/instruction JSON files.
data_root/
├── glioma/
│ ├── sub-001/
│ └── ...
├── meningioma/
├── ependymoma/
└── ...
For BraTS-related samples (once added by the user), the expected path format is: glioma/16_t1_t2_t1ce_flair_mask/...
📄 Files Included
1. Benchmark Data (Benchmark/ folder) 🏆
Benchmark/Benchmark_VQA_Open.json: Open-ended VQA pairs for evaluation.Benchmark/Benchmark_VQA_Closed.json: Closed-ended VQA pairs for evaluation.
2. Training Data (training/ folder) 🚀
training/train_cot_closed.jsonl: Chain-of-Thought data for closed-ended tasks.training/train_no_cot_closed.jsonl: Standard closed-ended QA pairs.training/train_open.jsonl: Open-ended instruction tuning data.
3. Metadata Files (metadata/ folder) ℹ️
metadata/all_brain_tumor_metadata.json: Standard dataset-level index.metadata/all_brain_tumor_metadata_rich.json: Extended metadata with radiological attributes.metadata/Dataset_For_LLM_Step1.json: Pre-processed data used for inference steps.
4. Image Data (images/ folder) 🖼️
images/Dataset.zip: Full dataset archive.images/Benchmark_Images.zip: Evaluation set images only.
📜 License & Ethics
License Aggregated Data: The images are aggregated from multiple publicly available medical imaging datasets. Each subset follows the license of its original source.
BraTS Data: Subject to the license and terms provided by the BraTS organizers.
Usage: This collection is provided strictly for research and non-commercial purposes.
Ethical Statement This dataset contains de-identified medical imaging data.
It is intended solely for research in Medical AI and Multimodal Reasoning.
⛔ It must NOT be used for clinical decision-making or diagnostic purposes.
📝 Citation
If you use the MM-NeuroOnco dataset or benchmark, please cite our paper
🔐 Access Terms and Conditions
By requesting access to this dataset, you agree to the following terms:
Research Only: Use the data strictly for academic and non-commercial research purposes.
Compliance: Comply with the original licenses of all source datasets.
No Redistribution: Do not redistribute any part of this dataset.
No Clinical Use: Do not use the data for clinical or diagnostic decision-making.
Citation: Properly cite MM-NeuroOnco in any derived work.
The authors reserve the right to deny access requests that do not comply with these conditions.
- Downloads last month
- 55