(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: The reconstruction of 4D CT images involves the binning of projection data to each temporal phase, and may suffer from deteriorated image quality due to inaccurate or uneven binning, e.g., under the non-periodic breathing, limited by the number of phases to bin. Cine CT can reconstruct a large number of phases so that the binning can be more accurate. However, the image reconstruction becomes more severely ill-posed due to smaller number of projections per phase. This work will develop novel deep learning regularized image reconstruction method for accurate cine CT reconstruction.
Methods: The new method parameterizes Douglas-Rachford iteration by neural networks in its natural manner of fixed-point iterations. The parameterized Douglas-Rachford fixed-point iterations of sparsity optimization (SOFPI-DR-Net) is an end-to-end mappings in which the inversions are parameterized by the physical forward operator and learned sparse transform, and the image priors are parameterized by U-nets for all phases jointly due to the observation that the phase images have highly similar topological structure. In our SOFPIDR-Net, the forward mapping and backward gradient of inversion are analytically provided by solving a linear system. Moreover, as a multilevel training strategy, we train the SOFPI-DR-Net entirely but regularize the intermediate reconstructed images in each stage to the solution space with increasing scale.
Results: The new DL methods are compared with conventional total-variation-regularized iterative methods. Both quantitative and visual results suggest that our SOFPI-DR-Net significantly outperformed classical 4D reconstruction methods, especially with CNN parameterized inversion methods.
Conclusions: We have developed a new image reconstruction method SOFIP-DR-Net for cine CT image reconstruction, with significantly improved image quality from classical 4D iterative methods.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
