(Tuesday, 7/31/2018) 1:45 PM - 3:45 PM
Room: Room 207
The essential components of CT imaging - data acquisition and image reconstruction - are typically performed on the same CT scanner using different features of the scanner software (i.e. the scan and recon functions). This coupling of the two functions is necessary given the impact of scanner geometry and the proprietary format of projection data. In the past decade, the introduction of iterative reconstruction (IR) algorithms has only cemented this coupling. As a result, state-of-the-art reconstruction algorithms are available on premium scanners or via software upgrades. Additionally, all IR algorithms consist of trade-offs, either in terms of computational speed, resolution, and/or noise texture. Thus the benefits in terms of data acquisition (e.g. speed) must be weighed against the reconstruction of a scanner. Here, we propose a method for cutting the cord between acquisition and reconstruction to obtain the maximum benefits of each. The method presented relies on the popular deep learning method using residual learning of a convolutional neural network (CNN). To train the network, FBP and IR reconstructed pediatric chest CT images of are divided into several 64×64 patches. Then 10,000 FBP-IR pair patches are randomly selected as the training data. We train 400 epochs for the proposed network. The code is implemented in Matlab (R2016a) environment running on a PC with AMD FX(tm)-6300 CPU and an Nvidia GeForce GTX 970 GPU. In addition, MatConvNet toolbox is used for the network training. The CNN reconstruction was then applied to a different dataset of FBP images from the same scanner and an older (10 year old) scanner capable of producing only FBP images to assess feasibility of the approach.
Authors:
Robert D. MacDougall, Yanbo Zhang
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
