Algorithmically Accelerated Iterative Reconstruction for Fast and Cost-Effective CT-based Explosive Detection Equipment

Accurate, realtime detection of explosives is a demanding application. High detection accuracy with a low false positive rate is desired. Noise and artifacts in reconstructed images, especially in the presence of metal, degrade the ability of detection algorithms to identify object's shape, volume, and composition. Model-based iterative reconstruction (MBIR) has been demonstrated to improve image quality over conventional direct reconstruction techniques - improving image resolution while suppressing noise and artifacts. The drawback is the significant increase of computation required for image formation, leading to an algorithm that is infeasible: either the reconstruction is too slow, or the hardware required for the desired throughput is too expensive. We will first establish a baseline iterative algorithm matching published state of the art image quality improvements. We will then reduce its computational demands 60 fold via algorithmic speedup. Cornerstone to this effort are the InstaRecon fast hierarchical operators, which reduce the computational complexity of the main computational burden of MBIR. Additional sources of algorithmic acceleration include variable splitting techniques for improved convergence rate, and approximate gradients. We will assess the combination of these algorithmic accelerations with hardware acceleration such as GPUs in the final technical report. Computation is a limiting factor in bringing iterative reconstruction to the market. Only so much hardware acceleration can be used without making cost a prohibitive factor. The algorithmic acceleration proposed here is an essential component of an iterative reconstruction system that can run at the required throughput on a modest hardware platform, making commercial deployment economically feasible.