Three-dimensional tomosynthetic image restoration for brachytherapy source localization

Tomosynthetic image reconstruction allows for the production of a virtually infinite number of slices from a finite number of projection views of a subject. If the reconstructed image volume is viewed in toto and the three-dimensional (3D) impulse response is known with a high degree of accuracy, then it is possible to solve the inverse problem using canonical image restoration methods by extension to all three dimensions in both the spatial and frequency domains.; This dissertation presents both modified direct and iterative restoration methods for solving the inverse tomosynthetic imaging problem in (3D) The significant blur artifact that is common to tomosynthetic image reconstructions is deconvolved by operating on the entire volume at once. The (3D) impulse response is computed analytically using a fiducial reference schema as realized in a robust, self-calibrating solution to generalized tomosynthesis. This imaging system is seeing-limited and may be regarded as a communication channel, thus allowing for its characterization in terms of Shannon's statistical information theory.; The relevant clinical application addressed in this dissertation is (3D) imaging for brachytherapy source localization. Conventional localization schemes for brachytherapy seed implants using orthogonal or stereoscopic projection radiographs suffer from scaling distortions and poor visibility of implanted seeds, resulting in compromised source tracking and dosimetric inaccuracy. With a well chosen projection sampling scheme, (3D) image reconstruction using the approach presented in this work increases the seed localization accuracy via 3D segmentation, leading to greater dosimetric precision.; Computer simulations, 3D impulse response, and 3D modulation transfer function (MTF) analyses were used to characterize the performance of the restoration methods. The restored images were then measured for their efficacy in the brachytherapy source localization task. Results show that the iterative conjugate gradient least squares (CGLS) algorithm yields the best 3D image restoration based upon the task criteria. Full 3D deconvolution is accomplished in clinically acceptable times with the Wiener and preconditioned CGLS algorithms resulting in reconstructions with significantly less tomosynthetic blur artifact. 3D source localization is then accomplished with greater accuracy vis-à-vis conventional plane film methods.