How to Pseudo-CT: A Comparative Review of Deep Convolutional Neural Network Architectures for CT Synthesis

This paper provides an overview of the different deep convolutional neural network (DCNNs) architectures that have been investigated in past years for generation synthetic computed tomography (CT) or pseudo-CT from magnetic resonance (MR). The U-net, Atrous-net and Residual-net were analyzed, implemented compared. Each was using 2D filters 3D with slices patches respectively as inputs. Two datasets used training evaluation. first one is composed by pairs T1-weighted MR Low-dose CT images head 19 healthy women. second database contains dual echo Dixon-VIBE pelvis 13 colorectal 6 prostate cancer patients. Bone structures target anatomy key choosing right learning approach. work a explanation order to know which DCNN fits better each medical application. According this study, U-net architecture would be best option generate pseudo-CTs while most accurate results anatomy.