Dual-energy computed tomography (DECT) technique has been widely investigated for decades since it generates more informative and accurate images than single energy CT. In DECT, two measured transmission data acquired at different energies are utilized to reconstruct numerous types of images. Our dual-energy joint statistic iterative image reconstruction algorithm, dual-energy alternating minimization (DEAM), has shown high accuracy in estimating proton stopping power mappings for proton treatment plan from axial experimental 3mm-collimated data. In this work, we try to transition the DEAM algorithm to a more clinical environment while retaining the same performance as in phantom experiments. In order to reconstruct the clinical data with enormous size, the 3D GPU-based helical DEAM is implemented, showing comparable performance to the 2D axial DEAM. Furthermore, DEAM is incorporated with image registration mappings to reduce patient movement artifact caused by sequential data acquisition, reducing the estimation uncertainty in mismatch regions from 10% to 1%. Moreover, two preliminary metal artifact reduction (MAR) methods are introduced to reduce the medical implant artifacts by down-weighting the data in the metal trace. Both methods show their potential for MAR. A set of tests is designed to find the principal cause for metal artifacts in DEAM results. The goal is to develop a MAR technique that reduces the uncertainty in the metal trace to 1%. Finally, machine learning approaches are investigated to reduce elapsed time of the DEAM towards the clinically acceptable time. The proposed recursive residual learning network with the reference update achieves 4X speedup compared to DEAM, with respect to time. Moreover, the network output could be used as the initial condition of DEAM if higher accuracy is desired. However, more data will be introduced to the training process to address the overfitting issue, and the elapsed time must be further reduced.

Dual-energy computed tomography (DECT) technique has been widely investigated for decades since it generates more informative and accurate images than single energy CT. In DECT, two measured transmission data acquired at different energies are utilized to reconstruct numerous types of images. Our dual-energy joint statistic iterative image reconstruction algorithm, dual-energy alternating minimization (DEAM), has shown high accuracy in estimating proton stopping power mappings for proton treatment plan from axial experimental 3mm-collimated data. In this work, we try to transition the DEAM algorithm to a more clinical environment while retaining the same performance as in phantom experiments. In order to reconstruct the clinical data with enormous size, the 3D GPU-based helical DEAM is implemented, showing comparable performance to the 2D axial DEAM. Furthermore, DEAM is incorporated with image registration mappings to reduce patient movement artifact caused by sequential data acquisition, reducing the estimation uncertainty in mismatch regions from 10% to 1%. Moreover, two preliminary metal artifact reduction (MAR) methods are introduced to reduce the medical implant artifacts by down-weighting the data in the metal trace. Both methods show their potential for MAR. A set of tests is designed to find the principal cause for metal artifacts in DEAM results. The goal is to develop a MAR technique that reduces the uncertainty in the metal trace to 1%. Finally, machine learning approaches are investigated to reduce elapsed time of the DEAM towards the clinically acceptable time. The proposed recursive residual learning network with the reference update achieves 4X speedup compared to DEAM, with respect to time. Moreover, the network output could be used as the initial condition of DEAM if higher accuracy is desired. However, more data will be introduced to the training process to address the overfitting issue, and the elapsed time must be further reduced.