(Sunday, 7/29/2018) 3:00 PM - 6:00 PM
Room: Exhibit Hall
Purpose: In four-dimensional (4D) cone-beam computed tomography (CBCT), there is a spatio-temporal tradeoff that currently limits the accuracy and quality of images. The aim of this study is to combine compressed sensing with generative adversarial learning for high quality 4D CBCT image reconstruction from a limited number of in-phase projections and prior CT images.
Methods: The collected 4D CBCT data are first applied to the conventional filtered backprojection (FBP) resulting in the images with the motion artifacts. A generative adversarial network, which is trained with the blurry 4D CBCT images and the sharp prior CT images, generates CT-like images from the blurry 4D CBCT images. Then, the 4D CBCT projections are grouped corresponding to the phases.The in-phase projection data and the generated images are plugged into a compressed sensing (CS) method based on least-square criterion regularized by total-variation and sparse image difference.
Results: The performance of the proposed algorithm is demonstrated through a series of simulation studies and experiments, and the results are compared to those of previously implemented compressed sensing techniques as well as conventional FBP-based 4D CBCT results. With such small number of in-phase projections, the conventional FBP failed to yield meaningful 4D CBCT images,and exiting CS techniques were not able to recover sharp edges.
Conclusion: The proposed method significantly reduces the gantry turns to achieve the high quality images compared to the 4D CBCT imaging based on the conventional FBP technique and the existing CS techniques.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
