基于双域迭代残差网络的双能CT图像材料分解研究

Research on dual⁃energy CT image material decomposition based on dual⁃domain iterative residual networks

针对在图像域将深度学习与迭代重建算法结合的深度迭代残差网络分解得到的材料图像受到噪声和伪影的影响,提出将迭代残差网络扩展到双域,即基于双域迭代残差网络的双能CT图像材料分解方法。该方法集成了两个并行交互的子网络CNN,同时在图像域和投影域进行材料分解操作,通过CNN直接向该网络提供投影数据,在不同域之间进行信息传递和融合,使用CNN丰富双域模型的数据保真度。实验结果表明,双域迭代残差网络相比于只在图像域进行材料分解能够更好地抑制噪声和伪影,提高图像质量和分解精度,做到细节保护。

In the image domain,the material image obtained by decomposing the deep iterative residual network which combines deep learning and iterative reconstruction algorithm is affected by noise and artifacts.To address this issue,the iterative residual network is extended to the dual⁃domain,which is a dual⁃energy CT image material decomposition method based on the dual⁃domain iterative residual network.This method integrates two parallel and interactive sub⁃networks CNN,and performs material decomposition operations in the image domain and the projection domain at the same time.The projection data is directly provided to the CNN,and the information is transmitted and fused between different domains.The CNN is used to enrich the data fidelity of the dual⁃domain model.Experimental results show that the dual⁃domain iterative residual network can better suppress noise and artifacts,improve image quality and decomposition accuracy,and sufficiently protect details than that perform image material decomposition only in the image domain.