ABSTRACT

GPU Execution Speed52 4.2Parameter Optimization Framework 52

4.2.1Assessing Reconstruction Quality Automatically-Image Quality Metrics52

4.2.2Optimization Algorithm-e Ant Colony Simulation 53 4.2.3Multi-Objective Optimization-Setting Fixed Objectives 55 4.2.4Multi-Objective Optimization-Assessing the

Entire Pareto Frontier56 4.2.5Learning from Experience-Applying the Optimized

Parameters for Novel Reconstructions 59 4.3 Conclusions 60 References 61

Iterative cone-beam CT algorithms have become increasingly popular in recent years. ey have been found useful when the projections are limited in number, irregularly spaced, or noisy. ese conditions arise, for example, in low-dose CT [4,8,11,15], where one reduces the x-ray beam intensity or tube current per projection and/or cuts down on the total number of projections to lessen the radiation dose to the patient. Low-dose CT has become a mission of great importance in recent years due to reports that the x-ray energy imposed onto patients during a CT scan can cause cancer. But there are also other imaging scenarios that can lead to sparse x-ray data, such as lack of time for acquisition or reduced angular access. In any of these cases, analytical techniques, such as the Feldkamp cone-beam algorithm [5], tend to produce reconstructions with strong streak and noise artifacts, which make reading these images for diagnostics dicult. Iterative reconstruction methods, on the other hand, in particular when augmented with some form of regularization, such as total variation minimization (TVM) [10,11] or nonlocal means (NLM) ltering [1,7,17], can overcome these challenges. Based on numerical optimization, they produce reconstructions that best t the data as well as some prior expectation of the object, formulated in the regularization function.