FEAT – FAST EFFICIENT ACCURATE TOMOGRAPHY

Achieving routine sub-100-nm resolution with x-ray nano-focus sources in the lab is challenging for two reasons: (i) the image contains phase-contrast features. These need to be treated properly, to achieve the desired resolution. (ii) The limited photon flux requires long exposures that are limited by the total available scanning-time and thermal drifts. Therefore, high-resolution tomography can only be achieved when the angular distribution of projections is as sparse as possible, decreasing total scanning time. In addition, smart choices of projection angles (unlike uniform distribution) and direction dependent exposition times could further lower the total scanning time. On the other hand, reducing projection angles is well known to corrupt the reconstruction by artifacts preventing a high-resolution reconstruction. 

Recent advances in Deep Learning have shown that sparse angle induced artifacts can be solved in the medical context where the availability of dedicated homogeneous CT training data is of no concern. Unlike in the general non-destructive testing scenario where the sample-to-sample variance can be as large as possible, the medical CT data shows patient images that resemble each other tremendously more. In addition, it has been indicated that reinforcement learning could be used to optimize angular and dose distribution. The project investigates both approaches for increasing tomographic resolution.