View optimization in Computed Tomography

Abstract

Many emerging CT imaging applications must operate with data from sparse projection angles, on the order of 10 views, such as first responder scenarios, dynamic imaging of detonation events, and flash imaging of materials under extreme conditions in high energy-density science. Few-view imaging is a data undersampling problem, that degrades tomographic reconstructions with streaks, shadows, and other aliasing artifacts, due to ill-conditioning. In few- view CT, the specific set of projection angles can strongly impact reconstruction quality, and the optimal angles depend on the object being scanned. Despite a surge of recent research, there is much to explore in this problem. We present an object-aware performance prediction framework for few-view projection datasets based on a matrix decomposition of the discretized forward model augmented with expected object gradient information. The method yields a quantitative score for a candidate set of view angles that can underpin a view selection recommendation algorithm. A key feature is that it can be applied at multiple stages of acquisition: a priori, before any radiographs are taken, or adaptively, after some radiographs are collected, to determine the next best view. We demonstrate the approach using simulated CT data with 5 view and 15 view configurations. Preliminary results are encouraging; images reconstructed using a standard statistical algorithm show that reconstruction accuracy agrees with our performance score. Reconstruction accuracy was evaluated against a known ground truth using an information theoretic score. Continued testing with complex objects and multiple geometries is necessary.