Photoacoustic Image Reconstruction Based On Group Sparse Dictionary Learning

Photoacoustic imaging is a non-invasive imaging technology based on photoacoustic effect.It breaks through the traditional depth limit of ballistic optical imaging and the resolution limit of diffuse optical imaging,and combines the advantages of high contrast of pure optical imaging and high spatial resolution of pure acoustic imaging.The structural imaging and functional imaging of biological tissues are realized,which provides a scientific and feasible new strategy for pre-clinical and early diagnosis of clinical diseases.The essence of photoacoustic image reconstruction is to invert the acoustic signal received by the ultrasonic transducer into an initial sound pressure distribution image of the biological tissue.Compared with the reconstruction methods based on analytical solutions,the model-based photoacoustic image reconstruction algorithms have many advantages,such as being free from the spatial distribution of ultrasonic transducers and good anti-noise performance,which have been widely used.In the model-based photoacoustic image reconstruction algorithm,there are two main implementation processes:forward model discretization and iterative solution of the target image.The forward model discretization method is usually based on interpolation,which has attracted much attention because of its simplicity and convenience.However,the traditional interpolation method only uses the information of a small number of adjacent pixels when calculating the pixel value of discrete points,and the fitting ability of the surface is weak,which cannot protect the edge information well.At the same time,it is easy to cause artifacts in the calculation process,which affects the reconstruction effect.In addition,retrieving the initial sound pressure distribution in the tissue according to the forward model is a pathological inverse problem.To obtain the optimal approximate solution,many studies take advantage of the sparse characteristics of photoacoustic signals and use sparse coding as a constraint term to optimize the inversion process.However,the traditional sparse coding methods ignore the group or cluster structure between the atoms of the dictionary,and fail to make full use of the characteristic information of the signal.Focusing on the above issues,this thesis mainly carried out the following aspects of work:(1)In the model-based photoacoustic image reconstruction algorithm,selecting the appropriate model is the key step,and the model construction is mainly realized by the discretization of the integral part of the acoustic transmission equation.To better approximate the nonlinear characteristics of the data,this thesis proposes a high-order interpolation method to calculate the model matrix with bicubic interpolation and cubic B-spline interpolation method as the core,so that as many pixel source points in the discrete point neighborhood as possible participate in the fitting process.Firstly,the coverage Angle of the integral curve is uniformly discretized,and then the discrete points of the integral curve are located.Finally,the contribution weight of the pixel to the discrete integral is determined by the interpolation calculation of the discrete points and the discrete coverage Angle.The results of numerical simulation and in vivo mouse study show that the proposed method can obtain more accurate model matrix,and the reconstructed photoacoustic image can better retain the edge information,enhance the contrast,and improve the accuracy and stability of image reconstruction.(2)The model-based photoacoustic image reconstruction algorithm is an ill-posed inverse problem.To maintain the stability and accuracy of the reconstructed image,it needs to be constrained by regularization.Among them,sparsity regularization can better capture the characteristics of the data itself,but the traditional sparse representation theory ignores the group structure information between the atoms of the dictionary.To make full use of the structure of the data,this thesis proposes a framework for photoacoustic image reconstruction with a sparse dictionary as the sparse coding constraint.The group structure is divided according to the atomic structure similarity of the dictionary by clustering algorithm.At the same time,total variation regularization is combined to achieve the preservation of image edge,texture and other structural information.The results of numerical simulation and in vivo study in mice confirmed that the algorithm proposed in this thesis reduced the redundant information and noise of the reconstructed photoacoustic image,enriched the details of the image,and effectively improved the accuracy and contrast of the image.(3)Although the above group sparse dictionary considers the structured information between the atoms of the dictionary,it does not make full use of the geometric relationship between atoms in different groups.In this thesis,we further propose a graph regular term by using the atoms in the dictionary as nodes to construct a graph structure.The dictionary learned under this constraint can not only ensure group sparsity,but also maintain the geometric relationship between the atoms in the dictionary derived from the group structure,which further ensures the effectiveness of group sparse coding.A new regularization model for photoacoustic image reconstruction was constructed by combining the group sparse dictionary and total variation regularization based on graph regularization.The results of numerical simulation and in vivo study in mice show that the proposed method can make better use of the redundancy and sparsity characteristics of the signal,and has better noise suppression and image micro-structure preservation performance,so as to achieve accurate imaging of the initial sound pressure distribution.