Advancing the Frontiers of Deep Learning for Low-Dose 3D Cone-Beam Computed Tomography (CT) Reconstruction

Ander Biguri, Subhadip Mukherjee

Abstract: Image reconstruction from the X-ray attenuation measurement in computed tomography (CT) can be formulated as an inverse problem for recovering a function in the 3D space from its line integrals. Variational regularization with model-based regularizers has traditionally been a successful approach for CT reconstruction. Recently, researchers have shown that combining a model-driven approach with the power of machine learning and computation can improve image quality significantly while reducing the radiation dose. However, a systematic evaluation of deep learning-based approaches on a publicly available benchmark dataset has not been undertaken. The main objective of the challenge is to gain insights into the empirical performance of different data-driven approaches for clinical CT imaging. We utilize the LIDC-IDRI dataset [1] consisting of chest CT scans, and simulate clinical- and low-dose projections from these scans (ground truth). The algorithms are evaluated in terms of the mean-squared error (MSE) of reconstruction w.r.t. the reference ground-truth scans.