MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Reproducibility of CT Iterative Reconstruction Algorithms From Analytic Reconstitutions with Convolutional Neural Networks for Pediatric Brain Imaging

R MacDougall*, Y Zhang , H Yu , UMass Lowell, Lowell, MA

Presentations

(Sunday, 7/29/2018) 3:00 PM - 3:30 PM

Room: Exhibit Hall | Forum 9

Purpose: To reconstruct conventional filtered back projection (FBP) images with a trained convolutional neural network (CNN) to simulate commercial iterative reconstruction (IR) algorithms

Methods: A convolutional neural network was trained with pediatric brain CT images acquired on a Siemens FORCE and GE Optima 660 scanner. Image training data was reconstructed on the scanner with FBP and the IR algorithm used in clinical imaging. The image training data was oversampled by reconstructing at 0.6, 1.25, 2.5 and 5.0 mm for both FBP and IR. The training code was implemented in Matlab with the MatConvNet toolbox for network training.

Results: The CNN reconstruction shows marked improvement compared to the FBP images both in terms of noise power and noise texture, and matches the commercial IR target very closely (Figure 1). The difference between CNN and FBP is largely a noise map demonstrating the noise that was removed from the CNN reconstructions (Figure 2). More training data of different size and anatomy is expected to further improve the trained network.

Conclusion: It is possible to reproduce IR algorithms by training a convolutional neural network with clinical images. A potential application of this result is the ability to de-couple the acquisition and reconstruction stages of CT image acquisition such that each component can be optimized separately, something that has not been possible until now.

Keywords

CT, Reconstruction, Computer Vision

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email