| Description |
There are several methods for volumetric scene analysis. This research focuses on real-time optimization of the ray-driven projector for forward and backward projections in the application of medical x-ray imaging. One of the current methods used by Computed Tomography systems to produce 3D medical images is the mathematical calculation of the line integral to simulate X-ray physics combined with Jacobi iterations, termed the Joseph method of discrete projection[1]. The proposed method is to expand upon the ray-driven model and precompute a sparse matrix of voxel to pixel relations. The transfer of a computational barrier to memory is only recently viable in real-time due to GPU hardware advances. The parallelization potential means that the mathematical operations can be precomputed and, with the utilization of modern GPU memory transfer protocols, cycled through for Jacobi iterations that rely on a sparse matrix of operations. Thus, the theoretical potential of the proposed method is twofold. First, the mathematical kernel operations become matrix operations which is already a deeply studied optimization problem, and second, the precomputation of angular information allows for more complex data compression. The second half of the research focuses on the extent to which algebraic convergence can improve by utilizing a high-resolution input matrix while still operating under the constraints of real-time. And with a final exploration of application to higher resolution reconstructions. This paper focuses on determining an optimal balance between algebraic reconstruction convergence, image quality, and time for a precomputed matrix. |
