基于卷积稀疏自编码的稀疏CT重建算法研究

Research on Sparse CT Reconstruction Algorithm Based on Convolutional Sparse Self Coding

稀疏CT数据重建是医学影像技术领域的重要研究方向。传统算法通过对高维稀疏数据学习与分析直接重建出高维图像,重建效率受数据维度影响。为克服高维稀疏数据重建效率随数据维度增高而降低问题,论文提出基于非线性降维与低维空间数据重建的SDAE-CNN重建网络。该网络融合卷积神经网络与稀疏自动编码器,通过双向网络实现在保留几何结构条件下的非线性降维映射与反映射,加入噪声和稀疏约束项,提高网络泛化能力与稀疏性,并借助改进卷积神经网络扩展自动编码器深度,完成低维空间图像重建与非线性升、降维。通过仿真实验验证,该方法能有效完成编码孔径扫描下稀疏CT数据的重建,且重建图像峰值信噪比、与原图的结构相似性皆高于一般算法。

Sparse CT data reconstruction is an important research direction in the field of medical imaging technology.Tradi⁃tional algorithms directly reconstruct high-dimensional images by learning and analyzing high-dimensional sparse data,and the re⁃construction efficiency is affected by data dimensions.In order to overcome the problem that the reconstruction efficiency of high-di⁃mensional sparse data decreases with the increase of data dimensions,this paper proposes a SDAE-CNN reconstruction network based on nonlinear dimensionality reduction and low dimensional spatial data reconstruction.The network integrates convolutional neural network and sparse automatic encoder,realizes nonlinear dimensionality reduction mapping and reflection with geometric structure reserved through bidirectional network,adds noise and sparse constraints,improves network generalization ability and sparsity,and expands the depth of automatic encoder with improved convolutional neural network to complete low dimensional spa⁃tial image reconstruction and nonlinear dimensionality increase and decrease.The simulation results show that this method can effec⁃tively reconstruct sparse CT data under coded aperture scanning,and the peak signal to noise ratio of the reconstructed image and the structural similarity with the original image are higher than the general algorithm.