Studies have shown that iterative algorithms outperform FDK, but their implementation is hindered by large demands on memory and computation time. "We noticed a big gap between research in image reconstruction and what is actually used in medical applications," explained Manuchehr Soleimani from the University of Bath. "Two important factors are that these algorithms are significantly slower than the standard FDK and that they are not accessible enough to non-experts."

To bridge this gap, Soleimani and colleagues from the University of Bath and CERN have developed the TIGRE (tomographic iterative GPU-based reconstruction) toolbox. Featuring a wide range of iterative algorithms, the toolbox is designed to be fast and easy to use, both for algorithm developers and medical end-users (Biomed. Phys. Eng. Express 2 055010).

What's in the box?

The main building blocks of any iterative algorithm are the projection and back-projection operators. To address the computation demands of iterative approaches, the researchers used CUDA to optimize these two blocks for graphics processing units (GPUs), thereby exploiting the massive parallelization afforded by GPUs. "The only way of making the code fast is by programming specific GPU code in very low-level programming languages, such as CUDA," explained first author Ander Biguri. "However, these are hard to learn and use and programming in them is tedious."

So to make TIGRE more user-friendly, the researchers coded the actual algorithms in MATLAB, a high-level, programming language that's easy to use and intuitive. "We combined the best of both, by accelerating the most computer-expensive blocks in the GPU, but allowing users to just use them in MATLAB," added Biguri.

The toolbox currently incorporates algorithms from four reconstruction families: the standard FDK algorithm; a range of algorithms from the SART-type family (SIRT, OS-SART and SART); CGLS from the Krylov subspace family; and the total variation regularization methods ASD-POCS, OSC-TV, B-POCS-TV-β and SART-TV. By incorporating these "black box" algorithms, TIGRE makes it easy for researchers who are only interested in image quality to test different algorithms, without requiring knowledge of how they work.

Reconstruction results

To illustrate the functionality of their toolbox, the authors presented two reconstruction examples. First, they used three algorithms to reconstruct data obtained from the RANDO head phantom. Image cross-sections showed that FDK resulted in noise across the entire image and significant strike artefacts. The iterative methods OS-SART and CGLS created smoother images, removed most artefacts and exhibited clearer separation between tissues. Processing times were 20 s, 46 min 30 s (40 s per iteration) and 4 min 41 s (20 s per iteration) for FDK, OS-SART and CGLS, respectively.

The possibility of reconstructing full 3D images using a reduced radiation dose is an important feature for CBCT development, particularly for radiotherapy applications where the patient may require imaging at each treatment fraction. In a second test, the team used FDK, OS-SART and ASD-POCS to reconstruct data from just 20 projections of a 3D Shepp-Logan phantom. In this extreme case, the increased performance of the minimization algorithms over FDK was evident, especially for ASD-POCS. Processing times were all below one minute.

Look to the future

The researchers note while that the toolbox enables image reconstruction with complex iterative algorithms in just a few minutes, further improvements are possible. One possibility would be to implement the algorithms in C++/CUDA – which would improve computation time by up to 50%, but make it harder to add new algorithms. The authors consider that the advantages of a high-level programming language for new algorithms outweigh the benefits of doubling the speed, which is already reasonably good.

"We decided that the toolbox is reasonably fast now, almost as fast as it can be, and we can already reconstruct quite large images in a good time scale," explained Biguri. "So to improve the toolbox, we will focus on incorporating novel algorithms and methods that could help improve images for radiotherapy treatment planning. Our own particular interests are new algorithms for motion correction and few-projection tomography, for accurate tumour location and imaging dose reduction."

The TIGRE code, which was recently released to the community, is open source, allowing anyone to download, test, modify and improve it. "Several research groups have shown interest in TIGRE and some of them are already trying it," Soleimani told medicalphysicsweb. "Feedback has been quite positive, and it seems that the wide range of algorithms included in TIGRE and its geometric flexibility is very interesting to some research groups."

"Hopefully, people will engage with TIGRE," he added. "And we will have improvements coming not only from the current TIGRE team – the Engineering Tomography Lab at the University of Bath, and Steven Hancock and Manjit Dosanjh from CERN – but also from all of the research community."

Related stories

• Dose engine brings MC calculations to all
• CPU-based method cuts optimization time
• Iterative algorithm wins PMB citations prize
• Algorithm enables boost in CBCT quality