|
Filtered back projection (FBP) reconstruction is simple and computationally efficient and is used in many commercial CT (tomography) imaging products. However, higher Poisson noise levels or metal objects in the imaged area can lead to severe artifacts. Iterative reconstruction employs stochastic models for the imaging process and the characteristics of the medical images and can reduce Poisson noise and metal related artifacts. But it is computation-intensive and furthermore, its image models are relatively simple and cannot quite capture the highly complex nature of the medical images, leaving rooms for further improvement. Recent advances in neural networks and deep learning could offer potential solutions to overcome these two problems. Towards that end, most of the neural networks proposed so far for CT image reconstruction are feed-forward networks with CNN (convolutional neural network) and fully connected layers, attempting to learn the mapping from the projections or the FBP output to the reconstructed image. While these networks have demonstrated some promising reconstruction or post-processing results, their architectures are somewhat arbitrary and the question remains as to what would be a more principled way to find a good architecture, thereby further improving reconstruction results. One promising idea is to design the network structure based on signal processing principles such as MAP (maximum a posteriori) estimation and iterative optimization. In this work, we developed a novel RNN (recurrent neural network) based on an accelerated iterative MAP estimation algorithm. This network makes use of, rather than learn, the forward image model such that the learning can be focused on the image or prior model and acceleration. This has led to good reconstruction results where Poisson noise and metal artifacts are greatly reduced.
|
