Data-efficient methods for multi-channel x-ray CT reconstruction
The superiority of iterative reconstruction techniques over analytical ones is documented in a variety of CT imaging applications. However, iterative reconstruction techniques generally require a substantial increase in data processing time and required resources, slowing adoption of the state of the art. This problem is exacerbated in multi-channel CT reconstruction problems (e.g. dynamic and spectral CT) where the gap between the amount of data acquired and to be reconstructed is often exaggerated. To facilitate adoption of iterative reconstruction techniques, we propose methods which seek to improve data efficiency. Specifically, we define data-efficient methods as those that produce reliable results with respect to task-specific metrics while managing the total x-ray exposure, sampling time, computation time, and computational resources required. The development of such methods unifies several themes in CT research, including dose management, task-based optimization, clinically relevant timelines for data processing, and reconstruction from undersampled data. In this paper, we present complementary, data-efficient methods for cardiac CT reconstruction. We present a reconstruction algorithm which requires minimal parameter tuning to solve temporal reconstruction problems. The algorithm exploits spatially localized, voxel-centric, distance-driven projection and backprojection operators to promote computational efficiency. We validate the algorithm with numerical simulations, using the MOBY mouse phantom, and with in vivo mouse data. For the in vivo data, localized reconstruction reduces computation time by 50% and system RAM requirements by 60% relative to non-localized reconstruction. Following validation of our algorithm, we present preliminary in vivo temporal denoising results using a convolutional neural network which promise to further improve the fidelity and speed of our reconstructions in future work.
