Skip to content

Latest commit

 

History

History
62 lines (42 loc) · 2 KB

README.rst

File metadata and controls

62 lines (42 loc) · 2 KB
Raw

GPU-accelerated functions for DIPY

cuDIPY is a Python library providing GPU-based implementations of a subset of algorithms from the DIPY software library.

This repository is at an early stage of development and should be considered experimental. Longer-term, we plan to incorporate GPU support into DIPY itself rather than maintaining two libraries in parallel.

Documentation

The behavior of functions in this repository should match those in the main DIPY respository. See DIPY's Documentation for details of individual functions.

Requirements

The following requirements should be installed prior to installing cuDIPY.

numpy >= 1.17 cupy >= 9.0.0b1 (targeting CuPy 9.0 release in April 2021) nibabel >= 3.0.0 scipy >= 1.1

Installing cuDIPY

cuDIPY is not currently on PyPI, but can be installed from the repository using pip:

pip install git+https://github.com/dipy/cudipy.git

Available Functionality

This library implements what is currently a relatively small, but useful subset of DIPY. Currently this include some the following primary functionality:

  • Non-rigid registration via cudipy.align.SymmetricDiffeomorphicRegistration (SyN) using a normalized cross-correlation metric (CCMetric). The AffineMap, DiffeoMorphicMap and ScaleSpace classes have also been implemented.
  • Tissue segmentation using cudipy.segment.TissueClassifierHMRF.
  • Gibbs artifactd removal via cudipy.denoise.gibbs_removal
  • Noise estimation via cudipy.denoise.noise_estimate and cudipy.denoise.pca_noise_estimate.
  • Brain extraction via median_otsu

In general, the functions and classes operate in the same way as their DIPY counterparts, but take CuPy arrays as inputs rather than NumPy arrays. The functions have only been tested on NVIDIA GPUs, but some functionality may also work on AMD GPUs via CuPy's initial HIP/ROCm support.