CTOperators.jl

This package provides 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; each back projection is the exact adjoint of its forward projection.

Important

CTOperators.jl is in an early stage of development. 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.
• A distance-driven projector (De Man & Basu), which averages the image over the area each detector cell sees, for detector geometries.
• 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("CTOperators")

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.