Reconstruction Algorithm Research On Sparse Prior Combined With Region-shrinking For Fluorescence Molecular Tomography

According to transmission of near-infrared light in biological media, fluorescence molecular tomography uses reconstruction methods to reflect internal fluorescent targets. It observes biological processes and monitors drug therapy at the molecular level, providing a powerful help for preclinical diagnosis of disease and small animal researches. The reconstruction is highly ill-conditioned because of characteristics of high scattering and low absorption in the light transmission and insufficiency of collected fluorescence data. In this paper, taking fully into account a priori information of fluorescent signal sparsity, region-shrinking strategy is used to improve stability of target reconstruction. The greedy algorithm is accepted to achieve fast reconstruction. The sorted l₁ norm minimization is adopted to improve target resolution. The TV norm minimization is employed to achieve the effective shape recovery. The specific research work is as follows:1) The reconstruction via region-shrinking strategy and greedy algorithm. The reconstruction result of greedy algorithm is unstable. Hence, the region-shrinking strategy is combined with stagewise orthogonal matching pursuit algorithm to reconstruct targets several times on the basis of gradually shrinking permissible region. The experiments for single target and double targets verify accuracy and stability of this method. The results show that greedy algorithm based on region-shrinking enhances localization accuracy and quantitative distribution of reconstruction targets and effectively improves ability to distinguish multiple targets, which makes reconstruction results more stable.2) The reconstruction via sorted l₁ norm sparse regularization. The regularization parameter is usually a carefully selected constant for most of the regularization methods. In order to achieve adaptability, the reconstruction process is described as a sorted l₁ norm minimization problem. A sorted 1-one penalized estimation algorithm is adopted for sparse reconstruction. The regularization parameter is a nonincreasing sequence of nonnegative scalars. Combining region-shrinking strategy can improve reconstruction performance. The simulated and physical experiments study localization capability for single target, ability to distinguish multiple targets and stability of this method. The results indicate that the sorted l₁ norm regularization method has better localization accuracy, satisfactory fluorescence yield and significantly improves the computational efficiency.3) The reconstruction via TV norm sparse regularization. Considering that TV norm minimization better maintains image edge details in image restoration, we study a TV norm minimization method realizing shape reconstruction. A total variation minimization by augmented Lagrangian and alternating direction algorithms is employed to recover shape structure of targets. The alternating direction method is utilized to minimize augmented Lagrangian function. Combining region-shrinking strategy can improve algorithm stability. The experiments for two-dimensional and three-dimensional validate feasibility in shape reconstruction of this method. The results demonstrate that TV norm regularization method not only accurately and efficiently reconstructs fluorescence targets, but also stably restores targets with different geometry shape characteristic.