FastDiag Toolkit
Efficient Medical Imaging Toolkit
Motivation
Medical imaging AI requires reliable data pipelines and consistent spatial semantics. Established Python libraries such as pydicom, SimpleITK, and nibabel remain focused on traditional imaging workstations, and their designs predate today’s requirements for concurrent loading, compute orchestration, and large-scale data management. When building a new generation of imaging platforms, issues such as fragmented datasets, missing metadata, and ambiguous coordinate systems repeatedly slow down development.
FastDiag Toolkit was created to close these gaps. It provides modern building blocks for AI workflows so that training, inference, and clinical systems can exchange data without losing fidelity or performance.
Core Components
DiCube: Medical Image Sequence Storage
DiCube stores an entire DICOM series in a .dcbs single-file container, optimized for high-concurrency I/O and cold-data archiving.
- Sequential reads minimize random seeks on mechanical drives
- HTJ2K lossless compression typically yields a 3× size reduction and a 5× read speedup
- Shared and slice-level metadata are separated and indexed
- Full round-trip conversion to standard DICOM remains available
MedMask: High-Efficiency Segmentation Masks
MedMask introduces the .msk format for segmentation masks, addressing the limitations of nii.gz in compression, semantic richness, and loading throughput.
- Zstandard compression commonly exceeds a 50× reduction
- Single-pass reading improves performance by roughly 16×
- Built-in semantic mapping supports multi-organ groupings
- Overlapping masks and multi-view alignment are first-class concepts
SpaceTransformer: Spatial Geometry Abstraction
SpaceTransformer models shape, spacing, origin, and orientation as a unified Space abstraction. It separates transformation planning from execution to guarantee accurate 3D geometric operations.
- Space-level planning enables chained transformations with just-in-time sampling during execution
- Images, masks, and landmarks remain aligned, avoiding
align_cornersambiguities - PyTorch backends offer GPU acceleration and integrate with common deep learning stacks
Installation
# Install the core libraries
pip install dicube
pip install medmask
pip install spacetransformer-core
pip install spacetransformer-torch # Optional: GPU accelerationGetting Started
Use the navigation on the left to explore design rationales, implementation details, and worked examples for each component. Reading the sections sequentially provides the quickest path to understanding how FastDiag Toolkit streamlines storage, compression, and spatial transformation in medical imaging workflows.
Reach out via the respective GitHub repositories for support or to contribute.