The ability to look inside an object without destroying it is useful for many applications in science, industry, and medicine. In tomographic imaging, projection images of the object are acquired along diﬀerent directions using some kind of penetrating beam. From these projection images, the 3D interior can be computed using tomographic reconstruction methods.
Tomography is the technique behind many 3D imaging devices and techniques such as medical CT scanners, micro-CT laboratory setups, 3D imaging using electron microscopes, and synchrotron tomography.
Reconstruction is typically performed oﬄine, after the scan has completed. If instead the reconstruction can be performed online and in real time, then the insight gained from the 3D reconstruction of the object can be used to immediately steer the experiment. Acquisition parameters such as source and detector positioning could be optimized based on the internal structure of the specific object, and dynamic processes in the imaged object could be followed as they occur. However, the runtime of conventional reconstruction algorithms is much longer than the time it takes to perform the scan, making it impossible to visualize the imaged object in real time.
In his dissertation, Jan-Willem Buurlage introduces various techniques that significantly reduce the time it takes to run conventional tomographic reconstruction algorithms without aﬀecting image quality. The resulting methods and software implementations put reconstruction times in the same ballpark as the time it takes to do a tomographic scan, enabling real-time tomographic reconstruction.
One of the newly proposed techniques is quasi-3D reconstruction. Instead of reconstructing the full 3D image, Buurlage shows that it is possible to reconstruct arbitrary oblique slices through the object at a fraction of the cost. By combining multiple slices, and allowing them to be chosen on-the-fly without reprocessing the measurement data, the operator of the scan has the illusion of having access to a full reconstructed 3D volume. Slices are reconstructed in a fraction of a second, enabling the real-time visualization of an object's interior. This method has already been successfully applied in practice at the FleX-ray lab of CWI, for electron tomography in a collaboration with EMAT in Antwerp, and for synchrotron tomography in a collaboration with the TOMCAT beamline of the Swiss Light Source.
Parallel reconstruction methods
Buurlage also introduces new algorithms to compute eﬃcient data partitionings, that allow advanced reconstruction methods to scale to dozens of GPUs. Together with the novel parallelization schemes and more eﬃcient communication data structures that are also introduced in the dissertation, these partitionings have the potential of greatly accelerating more costly algorithms that generally lead to better image quality compared to quasi-3D reconstruction.
Buurlage will defend his PhD thesis Real-time Tomographic Reconstruction at Leiden University on July 1st, 2020. His research was performed in the Computational Imaging group of CWI, under the supervision of Joost Batenburg (CWI / Leiden) and Rob Bisseling (Utrecht University).