RSNA Scientific Program




Wednesday Morning * Room N133
(L-8)


1216 * 10:30 AM


Improving Accuracy of the Algebraic Reconstruction Technique by Efficient Projection Pixel Supersampling




K. Mueller, MS, Columbus, OH * R. Yagel, PhD * J. Cornhill, DPhil

PURPOSE: In CT, the algebraic reconstruction technique (ART) is often used for image reconstruction when the projections are noisy, sparse, or obtained at nonuniform angular spacing. Recently, ART was also shown to have advantages in PET. Reconstructions obtained with ART often exhibit considerable noise artifacts, which may both obscure important object detail and suggest false image features. While these artifacts can be reduced by applying sophisticated but inevitably imperfect filters for grid interpolation in both projection and reprojection phases, a considerable portion often remains. Suppression of these residual artifacts by relaxation methods may both eliminate fine object detail and increase computation time.
MATERIALS AND METHODS: This paper analyzes the reconstruction process of ART in light of sampling theory concepts.
RESULTS: Our analysis indicates that the remaining noise artifacts are predominantly attributable to aliasing caused by interpolation kernel imperfection. This aliasing can be greatly reduced by supersampling the projection image pixels in both projection phases. Since a new, efficient table-based backward-viewing method is used instead of the traditional raycasting approach, supersampling does not compromise computation time significantly.
CONCLUSION: Projection pixel supersampling leads to considerably improved reconstructions in only 2-3 iterations and allows the number of projections to be reduced to 40 or less.