BrainBox

A unified medical imaging dataset containing 134,201 brain MRI slices with bounding box annotations, designed for fine-tuning vision-language models (VLMs) on medical image localization tasks.

Dataset Summary

BrainBox consolidates multiple 2D and 3D medical imaging datasets into a standardized 2D format optimized for VLM training. Each image is paired with structured metadata and precise bounding box annotations for pathological findings.

• Total Images: 134,201 2D MRI slices
• Source Datasets: 8 publicly available medical imaging datasets
• Modalities: T1w, T1c, T2w, FLAIR, DWI, ADC, TRACE
• Orientations: Axial, Coronal, Sagittal
• Pathologies: Stroke (68,136), Glioma (57,964), Meningioma (4,168), Pituitary Adenoma (3,933)

Dataset Purpose

This dataset serves as a fine-tuning corpus for vision-language models to perform:

• Medical image interpretation and localization
• Pathology detection with spatial grounding
• Multimodal medical question answering
• Cross-modal medical image understanding

Data Processing

• 3D to 2D Conversion: 3D volumes converted to optimal 2D slices using intelligent slice selection
• Standardization: Unified schema across heterogeneous source datasets
• Quality Control: Verified image paths and validated bounding box annotations
• Format: JPG images (256×256) with JSON bounding box metadata

Usage

from datasets import load_dataset
import json

dataset = load_dataset("liamchalcroft/brainbox")

# Example: Load an image with its annotations
sample = dataset["train"][0]
image = sample["image"]
diagnosis = sample["diagnosis"] 
bboxes = json.loads(sample["bounding_boxes"])  # Parse JSON string

Schema

Field	Type	Description
image	Image	2D MRI slice (JPG format)
diagnosis	string	Primary medical diagnosis
pathology_subtype	string	Specific pathology classification
lesion_type	string	General lesion category (neoplasm/vascular_lesion)
bounding_boxes	string	JSON array of lesion bounding boxes
modality	string	MRI sequence type
orientation	string	Slice orientation
has_pathology	bool	Presence of pathological findings
subject_id	string	Anonymized subject identifier
Additional fields	-	Demographics, imaging parameters, anatomical locations

Bounding Box Format

[
  {
    "x_min": 128, "y_min": 62,
    "x_max": 139, "y_max": 66,
    "width": 11, "height": 4, "area": 44
  }
]

Applications

• VLM Fine-tuning: Train vision-language models for medical image understanding
• Object Detection: Develop pathology localization models
• Medical AI: Build diagnostic assistance systems
• Research: Cross-modal medical image analysis studies

Data Distribution

• Vascular Lesions: 68,136 images (acute, chronic, subacute stroke)
• Neoplasms: 66,065 images (glioma, meningioma, pituitary adenoma)
• Balanced Orientations: ~33% each (axial, coronal, sagittal)
• Multi-modal: 7 MRI sequence types

Limitations

• Converted to 2D format (original 3D context reduced)
• Heterogeneous source data quality
• Limited demographic metadata coverage
• Bounding boxes simplified from original segmentation masks

Citation

If you use this dataset, please cite the original source datasets. This is a processed compilation for VLM research purposes.

License

Individual source datasets retain their original licenses. Verify licensing terms for your specific use case.