Computed tomography (CT) has become a familiar medical imaging technology that is also used to image materials in laboratories and X-ray synchrotron facilities. As scientists push for ever-higher CT resolutions, the absorbed radiation dose increases dramatically. These high radiation doses can structurally damage samples, which limits follow-up testing and restricts studies of dynamic processes, such as cell metabolism or battery cycling.

Both hardware and computational solutions have been devised to ameliorate radiation damage. The hardware approach involves hardening the sample to reduce X-ray damage from high-dose scans. The most popular hardening method is cryogenic freezing (vitrification), which locks the sample’s structure into place.

Unfortunately, modifying X-ray nano-CT instruments with cryogenic capability is technically challenging. Obtaining better performance requires building dedicated non-flexible instruments. It also prevents observation of dynamic processes and severely limits the range of experimental pressures and temperatures.

By contrast, the computational approach utilizes low-dose scans to reduce damage. To counteract the poorer images resulting from low-dose scans, scientists attempt to pull more information with an advanced machine learning-based denoising method from the X-ray projections, which are two-dimensional patterns formed when an X-ray beam passes through a sample. Hundreds of two-dimensional projections, taken from different angles, are typically required for a 3-D tomographic image. A neural network learns through optimization the differences between a few low-dose/high-dose pairs. Using this map, convolutional neural network (CNN) could then enhance the remaining low-dose projections, thereby allowing the imaging of dose-sensitive brain tissue at the nanoscale.