Advanced Methodologies In DOT/DFT Image Reconstruction Using Parallel Computation Strategies

The optical imaging technique (such as diffuse optical tomography (DOT) and diffuse fluorescence tomography (DFT)) has been developed as an effective supplementary method in reporting the functional and molecular biological imformations, with comprehensive advantage of sensitivity, specificity and security. However, present DOT/DFT techniques meet up with severe defects in accurately modeling of photon propagation in complex structures, precisely reconstruction of optical parameters in large-sized tissues, and effectively use of full time-resoved measurements. The main purpose of this thesis is to find applicable advanced solutions to improve the image quality of DOT/DFT reconstruction, mainly on the base of parallel computation stategies.We have first developed a GPU-accelerated MC modeling for photon migration in arbitrarily heterogeneous turbid medium, to accurately describe the optically heterogeneous structure such as low scattering, high absorbing and void regions. And based on this forward calculation, a continuous-wave DOT and a time-domain DFT image reconstruction algorithms are proposed. Simulative and experimental results showed that the MC-based approach can retrieves on the physiological and pathological information with higher accuracy than the DE-based one.A steady-state combined fluorescence-optical tomography system is developed in our work, which employs a fiber-switch-based tandem series-to-parallel mode to achieve the tradeoff among the measuring time, probing sensitivity and cost effectiveness. For the applications of breast diagnosis, a high-contrast DFT-guided hemoglobin DOT reconstruction algorithm is proposed based on the GPU accelerated MC modeling of photon migration, expecting to effectively alleviate the ill-posedness of the standalone hemoglobin DOT.For the DOT reconstruction of large-sized or three-dimensional tissues, a full domain-decomposition scheme with combined multi-core CPU and multi-thread GPU parallelization strategy are employed to cope with the severe computation and storage burdens that come from repeated use of the forward and inverse solvers. Numerical and phantom experiments both demonstrate that the proposed method can effectively reduce the computation time and memory occupation for the large-sized problem, with improved quantitative performance of the images due to the use of whole-matrix-based method.To take advantage of the full time-resolved data in time-domain DFT, an efficient solution is proposed by approximating the time-domain radiance in Fourier-series and accelerating the frequency-domain diffusion equations of multiple sampling frequencies using a combined CPU-GPU accelerated strategy. Simulative and experimental results both showed that the proposed method can obtain decent time-domain DFT reconstruction images with significantly fewer computational resources than that using finite difference method.This thesis improves the reconstructed quality of DOT and DFT techniques by some effective tools, including Monte-Carlo photon migration model, multi-core CPU and multi-thread GPU acceleration. Results on MC simulator, self-designed steady-state DOT/DFT system and pre-existing time-correlated-single-photon-counting system all showed that, the proposed methods can achieve corresponding significant improvements on computation time, accuracy and spatial resolution.