Research On High Accuracy Registration Of Images Of Dual-energy CT Using Synchrotron Radiation

Dual-energy Subtraction Imaging can eliminate unnecessary backgrounds by subtracting two images of sample acquired from low and high energy separately, thereby enhancing element distribution information. As this, it is widely used in medical imaging and chemical analysis. Combined with dual-energy CT imaging, obtaining a three-dimensional elements information inside sample is possible. Compared to ordinary X-ray source, synchrotron radiation light source has an apparent advantage. But the limits of scale and the experimental conditions under synchrotron radiation, easily lead to image misalignment of dual-energy CT data.This paper proposes an image registration method of dual-energy CT based on mutual information, which uses mutual information as a measure of image similarity and multi-parameter optimization algorithm(GA and Powell algorithm) to search for optimal registration parameters in multi-dimensional space. This method can overcome the local minima problem of PV interpolation algorithm. Also, the high complexity of HPV interpolation can be solved using a multi-step sampling to ensure registration accuracy, meanwhile significantly reducing image registration time. The research experiment is conducted at Soft X-ray Spectroscopy Microscopy Beamline(BL08U) at Shanghai Synchrotron Radiation Facility(SSRF).Firstly STXM imaging of soil particle in micron size is needed. Then dual energy CT data is reconstructed by CT algorithm. Finally, image registration and subtraction of dual-energy CT is performed, and we get three-dimensional distribution of element. The main contributions are concluded below: 1. Verify that three-dimensional image registration algorithm based on mutual information is effective for dual energy CT data and can achieve sub-pixel accuracy. Based on this, comparison of mutual information curve positioning accuracy between PV interpolation and HPV and algorithm robustness with noise is investigated. And propose a criteria of sub-pixel registration accuracy and analyze results of simulated and experimental data. 2. Calculate mutual information through multi-step sampling which reduces the computational complexity of registration and ensures registration accuracy. This method can significantly improve program execution speed. 3. Collect soil particles of different soil layers. Choose iron-rich particles filtrated by a magnet. Then by micron dual-energy CT imaging, mutual information registration algorithm and subtraction, element distribution information is obtained. This lays a foundation for subsequent study of soil environmental sciences including analyzing information of Fe contained in soil particles and possible differences in soil layers.The second part of work is CUDA parallelization of partial modules of Equal Sloped Tomography(EST). Compared with FBP algorithm, EST can reduce radiation doses and number of the projection for reconstruction, so good image quality can be assured. But because partial data in pseudo polar coordinate is missed, an iterative process is required to make up this which leads to slow speed. CUDA(Compute Unified Device Architecture) is NVIDIA’s unified computing platform and application interfaces(API), allowing developers to use graphics processing unit(GPU) for general purpose parallel computing. So complete parallelization of two-dimensional modules of EST algorithm, including Fractional Fourier transform(Fr FT), pseudo polar coordinate Fourier transform(PPFFT) and inverse transform(IPPFFT).