radiobject 0.1.0

TileDB-backed data structure for radiology data at scale

RadiObject

What? A TileDB-backed data structure for radiology data at scale.

Why? NIfTI/DICOM must be read from local disk and don't support partial reads. TileDB enables cloud-native storage (S3), efficient partial reads, and hierarchical organization of multi-volume datasets.

Thoughts

Installation

pip install radiobject

Quick Start

from radiobject import RadiObject

# Create from NIfTI files using images dict (recommended)
radi = RadiObject.from_niftis(
    uri="./my-dataset",
    images={
        "CT": "./imagesTr/*.nii.gz",      # Glob pattern
        "seg": "./labelsTr",               # Directory path
    },
    validate_alignment=True,               # Ensure matching subjects across collections
    obs_meta=metadata_df,                  # Optional subject-level metadata
)

# Access data (pandas-like)
vol = radi.CT.iloc[0]            # First CT volume
data = vol[100:200, :, :]        # Partial read (only loads needed tiles)

# Filtering (returns views)
subset = radi.filter("age > 40")       # Query expression
subset = radi.head(10)                 # First 10 subjects
subset.materialize("./subset")         # Write to storage

Works with local paths or S3 URIs (s3://bucket/dataset).

How It Works

NIfTI requires decompressing entire volumes; TileDB reads only the tiles needed. This enables 200-660x faster partial reads. See benchmarks →

Sample Data

Download sample datasets for tutorials and testing:

# Install download dependencies
pip install radiobject[download]

# Download BraTS brain tumor data (for tutorials 00-04)
python scripts/download_dataset.py msd-brain-tumour

# List all available datasets
python scripts/download_dataset.py --list

Documentation

• Tutorials - Interactive notebooks
• Data Access - Ingestion, queries, filtering
• ML Integration - MONAI/TorchIO setup
• Design - Architecture decisions
• Benchmarks - Performance analysis
• Datasets - Available datasets and download instructions

License

MIT