Assessing the efficacy of tomographic reconstruction methods through physical quantification techniques

Conventionally the performance of computed tomography (CT) reconstruction algorithms is assessed by a voxel-to-voxel comparison between the object and the reconstructed volume, often using digital phantoms. However the real aim in the CT imaging community is not to develop a reconstruction algorithm to obtain the best-looking images, but one that allows us to extract the relevant information to a desired accuracy. Here, through various case studies, we quantify features of interest for the test object and use these as measures of the efficacy of the reconstructions. Where applicable, we compare the assessment technique against commonly used metrics to measure the quality of a reconstructed solution, and find that in most cases the popular metrics have no relation to the accuracy of the features we extract from a reconstruction. The assessment technique we demonstrate, which we refer to as physical quantification, is used to determine the shape, contacts and size of beads for a test dataset made available via the SophiaBeads Dataset Project. Using this image analysis approach a number of widely used reconstruction methods are evaluated. Our work shows that it is important to choose the optimal reconstruction strategy based on the features you want to quantify from the scan. For example, in our case we found that the shape of the beads could be measured using TV regularization with eight times fewer projections than the other methods, or that reconstructions obtained via many but noisy projections yield as accurate results as those obtained via less noisy but fewer projections.