CTOperators.jl

This package provides matched forward and back projectors for X-ray computed tomography that run multi-threaded on the CPU and on GPUs (CUDA, AMD, Apple) from the same code base.

Important

CTOperators.jl is in an early stage of development and not registered yet. Its interface may still change.

Features

• Projectors for iterative reconstruction: a projector is created once and applied many times; the back projection is the exact adjoint of the forward projection.
• Two ray-driven projectors: Siddon's method, which weights each voxel by the exact length of the ray inside it (in the incremental form of Jacobs et al.), and Joseph's method, which interpolates linearly between voxel centers.
• Generic geometries: arbitrary rays given by start and end points, and cone-beam (flat or curved detector) and parallel-beam scanners described view by view, with any source trajectory (circular, helical, ...), including objects that move during the scan. Geometries of Sinograms.jl work directly.
• One code base for CPU and GPU, written with KernelAbstractions.jl: data as Array, CuArray (CUDA.jl), ROCArray (AMDGPU.jl) or MtlArray (Metal.jl).
• Differentiable with respect to image and projection data (ChainRulesCore rules).
• Part of the JuliaImageRecon family: images are described by ImageGeoms.jl, reconstructions run with RegularizedLeastSquares.jl.

Example

using CTOperators

image = ImageGeom(; dims=(128, 128, 128), deltas=(1.0, 1.0, 1.0))
geometry = helical_cone_beam(; angles=range(0, 4pi, 360), pitch=64, z0=-64, dso=256,
                             dsd=512, nu=256, nv=128, du=1.0, dv=1.0)
op = SiddonProjector(image, geometry)        # Float32 on the CPU
x = rand(Float32, 128, 128, 128)
y = op * x                                   # projections, size (256, 128, 360)
z = adjoint(op) * y                          # back projection

On a GPU, create the projector with the GPU's array type, e.g. SiddonProjector(image, geometry; storage=CuArray), and pass data that lives on the GPU.

Installation

using Pkg
Pkg.add(url="https://github.com/JuliaImageRecon/CTOperators.jl")

Benchmarks

benchmark/ compares CTOperators.jl with parallelproj and RadonKA.jl on the CPU and the GPU; its README describes how to run the benchmarks and lists the results.

Community Standards

This project is part of the Julia community and follows the Julia community standards.