Learning the Forward Operator in Photon-Counting Computed Tomography

Abstract: Computed Tomography (CT) is a non-invasive x-ray imaging method capable of reconstructing highly detailed cross-sectional interior maps of an object. CT is used in a range of medical applications such as detection of skeletal fractures, organ trauma and artery calcification. Reconstructing CT images requires the use of a forward operator, which is essentially a simulation of the scanning process. Photon-Counting CT is a rapidly developing alternative to conventional CT that promises higher spatial resolution, more accurate material separation and more robust reconstructions. A major difficulty in Photon-Counting CT is to model cross-talk between detectors. One way is to incorporate a wide point-spread function into the forward operator. Although this method works, it drastically slows down the reconstruction process.  In this thesis, we accelerate image reconstruction tasks for photon-counting CT by approximating the cross-talk component of the forward operator with a deep neural network, resulting in a learned forward operator. The learned operator reduces reconstruction error by an order of magnitude at the cost of a 20% increase in computation time, compared to ignoring cross-talk altogether. Furthermore, it generalises well to both unseen data and unseen detector settings. Our results indicate that a learned forward operator is a suitable way of approximating the forward operator in photon-counting CT.