The aim of this project is to develop a new x-ray micro computed tomography (micro-CT) technique that will allow performing high-resolution scans with significantly less radiation dose than conventional machines. In the latter, dose generally increases cubically with increasing the spatial resolution; therefore, high-resolution scans require high doses. This means that such scans are not applicable to many types of biomedical samples (e.g. live cell cultures, or small animals) due to their radiation sensitive nature. Our proposed approach will disrupt this relationship between spatial resolution and radiation dose, hence overcome this application restriction and facilitate a widespread use of x-ray micro-CT in bio-medicine. This will provide new options for validating biomedical research, e.g. by repeatedly scanning cell-bearing samples at high spatial resolution, which is not possible with current technology. In the long term, this will lead to new medical treatments, or more cost effective therapies for conditions that currently require expensive care.
Our proposed approach combines advanced x-ray engineering with machine learning. The engineering part will entail radically changing the experimental setup of x-ray micro-CT machines. Specifically, we will use a mask in the x-ray beam to create an array of narrow beamlets. This has a two-fold effect: first, spatial frequencies beyond the cut-off normally defined by the x-ray source and detector are inserted into the image formation process, leading to an increase in spatial resolution. Second, large parts of the sample are shielded from radiation, enabling significant dose savings. We will combine this modified setup with an acquisition scheme in which the sample is scanned along a cycloidal trajectory (achieved by simultaneously translating and rotating the sample). This is a key innovative aspect of this project; while the use of beamlets generally leads to incomplete data sets as some parts of the sample are not “seen”, a cycloidal acquisition allows controlling the precise locations of the available data, which provides a favourable configuration for restoring any missing information via mathematical methods. For the latter, we will exploit new developments in machine learning based on convolutional neural networks.
The breakthrough character of this project was demonstrated in two proof-of-principle studies. The first study has shown that our cycloidal acquisition approach can increase the spatial resolution in x-ray micro-CT images by at least a factor of three without requiring any increase in radiation dose. The second study has demonstrated that convolutional neural networks are well-suited for application in low-dose computed tomography and lead to a significantly improved image quality. We expect that, by combining both aspects, we will be able to increase the spatial resolution in micro-CT images much further than what we have already demonstrated, with no increase in radiation dose.