Parallelizable 3D statistical reconstruction for C-arm tomosynthesis system

Clinical diagnosis and security detection tasks increasingly require 3D information which is difficult or impossible to obtain from 2D (two dimensional) radiographs. As a 3D (three dimensional) radiographic and non-destructive imagine technique, digital tomosynthesis is especially fit for cases where 3D information is required while a complete projection data is not available. Nowadays, FBP (filtered back projection) is extensively used in industry for its fast speed and simplicity. However, it is hard to deal with situations where only a limited number of projections from constrained directions are available, or the SNR (signal to noises ratio) of the projections is low. In order to deal with noise and take into account a priori information of the object, a statistical image reconstruction method is described based on the acquisition model of X-ray projections. We formulate a ML (maximum likelihood) function for this model and develop an ordered-subsets iterative algorithm to estimate the unknown attenuation of the object. Simulations show that satisfied results can be obtained after 1 to 2 iterations, and after that there is no significant improvement of the image quality. An adaptive wiener filter is also applied to the reconstructed image to remove its noise. Some approximations to speed up the reconstruction computation are also considered. Applying this method to computer generated projections of a revised Shepp phantom and true projections from diagnostic radiographs of a patient's hand and mammography images yields reconstructions with impressive quality. Parallel programming is also implemented and tested. The quality of the reconstructed object is conserved, while the computation time is considerably reduced by almost the number of threads used.