Asymmetrical Dual-Cycle Adversarial Network for Material Decomposition and Synthesis of Dual-Energy CT Images

EasyChair Preprint 11596

Abstract

Dual-energy computed tomography (DECT) can identify the material properties with its excellent material quantitative analysis ability. However, the application of DECT is restricted by the problems of inaccuracy of energy spectrum estimation, non-linearity and inconsistency of imaging geometry, which will lead to the degradation of material distribution images. Hence, deep learning (DL)-based methods have become the state-of-the-art technique in DECT rely on its excellent feature recognition performance in the case of few spectrum prior. In this work, we propose an asymmetrical Dual-Cycle adversarial network (ADCNet) for both material decomposition and synthesis of dual-energy CT images, which has certain advantages in spectral CT multi-task parallel, improvement of image quality and radiation dose reduction. The experimental results show that the cycle framework achieves the adversarial learning of dual networks, and promotes the quality of generated images by introducing multiple mechanisms. Compared with the traditional DL-based methods, the proposed method has outstanding qualitative and quantitative indicators.

Keyphrases: Adversarial Learning, Asymmetrical Dual-Cycle structure, Spectral CT, material decomposition

@booklet{EasyChair:11596,
  author    = {Xinrui Zhang and Ailong Cai and Shaoyu Wang and Ningning Liang and Yizhong Wang and Junru Ren and Lei Li and Bin Yan},
  title     = {Asymmetrical Dual-Cycle Adversarial Network for Material Decomposition and Synthesis of Dual-Energy CT Images},
  howpublished = {EasyChair Preprint 11596},
  year      = {EasyChair, 2023}}