Study On Fast Iterative Reconstruction Algorithms For Digital Breast Tomosynthesis Imaging

Breast cancer is now the most common cancer in Chinese women,and screening is an important step in early detection of breast cancer.Digital mammography(DM)is a well-established imaging technique for early breast cancer screening.It has the advantages of high spatial resolution and low radiation dose.However,DM images only contain two-dimensional(2D)information of three-dimensional(3D)anatomy structures,for which the accuracy of breast cancer detection based on screening mammography is affected by the fact that the breast cancer is often obscured by overlapping tissues.Digital breast tomosynthesis(DBT)is an emerging imaging technique based on DM.In DBT,the volume image is reconstructed from 2D projections data acquired at a limited number of views over a limited angular range.DBT images can provide quasi-3D structural information of the breast,reduce tissue overlap and improve sensitivity for subtle mass lesions detecting.At present,DBT has become one of the important medical imaging modality in clinical application.At present,there are two categories of DBT reconstruction algorithms,namely analytical reconstruction and iterative reconstruction.The filter back projection algorithm(FBP)is the most classical analytical reconstruction algorithm.After obtaining the projection data,a special filter is designed to filter the projection.Finally,the filtered projection is backprojected to obtain the reconstructed image.We have to note that the filter largely depends on the image content and sampling geometry.Among the iterative reconstruction algorithms,the adaptive steepest descent-projection onto convex sets(POCS)proposed by Sidky et al.is the most widely used iterative reconstruction algorithm framework.Each iteration of the algorithm is divided into two steps:(1)reconstruction,using ART algorithm to reconstruct the initial image,reducing data errors;(2)denoising,using gradient descent method to solve regularization problems.Repeat until the end condition is satisfied.In this paper,we systematically review the composition of DBT system and its imaging theory,and deduce the classical DBT reconstruction algorithms.In order to solve the time-consuming problem in iterative reconstruction,the following two reconstruction algorithms are proposed:Firstly,a fast iterative reconstruction algorithm of DBT based on multi-GPU and distributed ADMM is proposed.In order to achieve fast iterative reconstruction of DBT images,we describe the problem of DBT reconstruction as a distributed optimization problem,constructing an energy functional which consists of fidelity terms and total variation terms,and is solved by using consensus Alternating Direction Method of Multipliers(ADMM)optimization algorithm based on multi-GPU hardware platform.In the algorithm,the projections are divided into N subsets,each GPU corresponds to a projection subset and performs the conjugate gradient to update the intermediate variables in parallel.After that,a parallel reduction among GPUs is conducted to compute the summation over all these intermediate variables and perform the image artifact processing.Then the updated image is broadcasted to all GPUs to ensure the same reconstructed volume data in each GPU.The experimental results show that images reconstructed by the proposed method show great contrast of the features and a speedup factor over 1.7 has been achieved compared with conventional methods.Secondly,a variant of the fast iterative shrinkage thresholding algorithm(FISTA)is proposed and applied to the iterative image reconstruction in DBT.Because of the burdensome computation of Lipschitz constant and the unstable update step size,the original gradient-descent step is replaced by a sub-problem that is solved by utilizing the ordered subset simultaneous algebraic reconstruction technique(OS-SART).In addition,due to the preconditioning matrix adopted in the OS-SART,a new weighted proximal problem is introduced and fast gradient projection(FGP)algorithm is employed to solve it.After that,the SLO minimization is used to further suppress noise and artifacts and obtain high quality image.The experimental results show that the improved algorithm,combining with SLO regularization,is superior in suppressing noise and artifacts while preserving structural edges.