Dual-energy X-ray imaging systems are widely used in medical imaging fields and industrial fields since they can provide additional material information as well as X-ray attenuation contrast of the imaged object. The material information greatly improves not only the detectability of imaging systems but also the accuracy of estimated physical properties. However, in dual-energy X-ray imaging, noise amplification, which is inevitably occurred due to the calculation between two images, and structural issues caused by dual-energy X-ray imaging approaches deteriorate image quality. Therefore, in this dissertation, we propose novel algorithms for various dual-energy X-ray imaging systems for the common purpose of improving the accuracy of material decomposition.
First, we propose a new clustering algorithm to reduce noise occurred in the cargo inspection system. In the dual-MV X-ray imaging system using a single linear accelerator (LINAC), low-energy images inherently have much higher noise level than high-energy images. To resolve this, we applied a relatively reliable cluster map, obtained from the high-energy image, to the low-energy image for effectively reducing the noise. The proposed clustering algorithm, which includes a regularizer of entropy form, automatically finds the optimal cluster number depending on the imaged object.
Second, we propose a reconstruction algorithm for the dual-energy CT (DECT) system using a multi-slit beam filter. The reconstruction algorithm improves the high-energy image quality by using the asymmetrically sharing sparsity of the relatively high-quality low-energy image. Also, we propose a one-shot material decomposition method that performs image reconstruction and material decomposition simultaneously. The one-shot material decomposition method is applicable to the systems where the shooting angles of low- and high-energy images do not match, such as kV-switching DECT, and effectively reduce beam-hardening artifacts through modeling of polychromatic X-rays.
Finally, we propose an energy-flexible network to reduce the burden of acquiring training data since energy pairs used in DECT are a lot varied according to the applications. In this study, we developed two training techniques that extend two given data spaces for energy flexibility.
We performed both simulations and real experiments in each dual-energy X-ray imaging system to verify the feasibility of the proposed algorithms, and the results show that all the algorithms successfully contribute to improving the accuracy of material decomposition. We expect that the proposed algorithms can be applicable not only to dual-energy X-ray imaging but also to other imaging fields where two images having mostly similar but one different characteristic are dealt with.