Abstract
We present the ASTRA (All Scale Tomographic Reconstruction Antwerp) toolbox: an open source, GPU-accelerated library for 3D image reconstruction in tomography. While most of the current software tools available for tomography offer only limited flexibility in the geometry of the experimental setup and the set of available reconstruction algorithms, the ASTRA toolbox provides a set of highly flexible building blocks that can be used to construct advanced reconstruction algorithms, effectively removing these limitations. We describe how the design of the ASTRA toolbox allows for full flexibility in specifying the geometry while still maintaining an efficient mapping onto the underlying hardware. The ASTRA toolbox comes with a MATLAB interface for easy user interaction and is available for both Windows and Linux.
Introduction
Tomographic reconstruction—the process of creating a 3D volume from a series of 2D projection images—is a significant computational bottleneck in fields like electron microscopy and synchrotron imaging. Traditional analytical methods like Filtered Backprojection (FBP) are fast but often yield poor results with noisy or limited data. Conversely, iterative algebraic methods produce higher quality images but are computationally intensive, requiring repeated simulations of the projection process.
Before the development of ASTRA, researchers were often forced to choose between rigid, high-speed "black-box" software and flexible but slow toolboxes. Real-world experiments frequently involve misaligned samples or non-standard trajectories that standard software cannot model. The ASTRA toolbox addresses this by bridging the gap between flexible algorithm prototyping and high-speed GPU execution.
Methods and Design Philosophy
The ASTRA toolbox is built on a "building block" philosophy. Rather than providing a single fixed workflow, it exposes highly optimized primitives—specifically Forward Projection and Back Projection—via NVIDIA CUDA for GPU acceleration.
Key technical innovations include:
- Geometric Flexibility: ASTRA allows users to specify the exact 3D position and orientation of the source and detector for every projection, accommodating experimental misalignments.
- Decoupled Resolution: Reconstruction voxel size is independent of detector pixel size, enabling super-resolution studies.
- Hybrid Interface: The core is written in C++/CUDA for performance, while a MATLAB interface allows researchers to implement complex mathematical optimization algorithms at production speeds.
- Primitives over Products: By exposing projection operators ($W$ and $W^T$) as functions, ASTRA enables the rapid implementation of advanced algorithms like Total Variation (TV) minimization.
Performance Evaluation
The authors evaluated ASTRA using three benchmark datasets (A, B, and C) typical of Electron Microscopy, with sizes up to $1424 \times 2024$ pixels, and true 3D cone-beam reconstructions with volumes up to $1024^3$ voxels.
Key Results
The performance gains are substantial. On the largest test dataset (Dataset C), ASTRA running on a mid-range GTX 680 GPU completed a SIRT iteration in just 6.67 seconds, significantly faster than previous GPU and CPU implementations.
The toolbox also demonstrates excellent multi-GPU scalability. For a massive $1024^3$ volume, increasing the hardware from a single NVIDIA Tesla M2090 to four GPUs reduced the iteration time from 26.13 seconds to 9.92 seconds.
Limitations and Future Work
Despite its strengths, ASTRA is currently limited to NVIDIA GPUs due to its CUDA dependency. Additionally, while multi-GPU support mitigates memory constraints, the massive size of modern 3D datasets (often exceeding $2048^3$) remains a challenge for GPU RAM. Future work will focus on leveraging this flexible foundation to develop advanced discrete tomography and sparsity-based algorithms that were previously computationally prohibitive.
Summary updated