Sparse-View CT Joint Reconstruction Strategy with Sparse Sampling Encoding Layer.

Abstract

Background: Sparse angular projection is an important way to reduce CT dose. It consists of two processes, sparse sampling, and image reconstruction based on sparse projection. Under the traditional reconstruction framework, the sparseness of the projection angle may cause a degradation effect in the reconstructed image. A series of machine learning methods for sparse angle CT reconstruction developed in recent years, especially deep learning methods, can effectively improve the reconstruction quality, however, these methods can only reconstruct CT images based on a certain sparse sampling scheme.

Objective: On the other words, they cannot search for an efficient sparse sampling scheme under a certain dose constraint automatically, which became the motivation to develop an end-to-end sparse angular CT reconstruction method.

Methods: In this work, we propose a sampling encoding layer for searching sparse sampling schemes and integrate it into a sparse reconstruction neural network model based on projection data. Meanwhile, a joint reconstruction strategy based on both the radon domain and image domain painting is also developed.

Results: Experiments based on public CT datasets demonstrate the effectiveness of the method.

Conclusion: The results show that the joint reconstruction network based on a sparse sampling coding layer has great application potential.