A new machine learning method to reconstruct images needed for neutron tomography can speed up the process, making it a more feasible imaging option.

Neutron tomography is computed tomography involving the production of three-dimensional images by detecting the absorption of neutrons produced by a neutron source. It’s a non-destructive way to investigate the inner structure of objects, and it can be more effective than more common X-rays. It is used in medical imaging of items such as bones and teeth; it’s also used in other areas, such as art history and plant physiology.

However, it’s time consuming, making the technique less practicable to use. Some scans can take several hours.

One way to decrease the scan time is to reduce the number of projections needed and rely on reconstruction of some of the images. However, while the analytic reconstruction by current algorithms is fast, the reconstructed images typically have imperfections, such as streaks.

Fedor Kondratenko

The researchers, from Italy and the United Kingdom, tested a recently introduced neural network filtered back projection (NN-FBP) method to optimize the time usage in neutron tomography experiments—it’s the first time this machine learning-based algorithm has been applied and tested to address the neutron tomography reconstruction problem.

They reconstructed a subset of images in a simulation study using a monoblock from a fusion energy device and trained the artificial neural network to mimic the reconstructed images. The NN-FBP method “significantly” outperformed conventional reconstruction algorithms with better image quality.

The researchers then conducted an experimental study with similar results. The NN-FBP method reduced the number of projections needed to one third, decreased acquisition and reconstruction times and outperformed conventional reconstruction methods used in neutron tomography.

“Deep learning and machine learning in general are promising and innovative approaches for image reconstruction…[T]he [neutron imaging] community should take into account new [machine learning] based-reconstruction theories and techniques,” the study authors stated.

The study was reported in Scientific Reports.

The researchers noted that this method may be of particular interest to those in medical and X-ray imaging, but that it can also have wider impact.

“NN-FBP can reliably reduce scan time, reconstruction time and data storage, providing high image quality .… hence, the NN-FBP method can be implemented with high computational efficiency at neutron imaging facilities for the broader applicability than regularized iterative reconstruction algorithms,” the study authors stated.