Task-based comparative study of iterative image reconstruction methods for limited-angle x-ray tomography

For tomography that has available only projection views from a limited angular span, such as is the case in an x-ray tomosynthesis system, the image reconstruction problem is ill-posed. Reconstruction methods play an important role in optimizing the image quality for human interpretation. In this work we compare three popular iterative image reconstruction methods that have been applied to digital tomosynthesis systems: the simultaneous algebraic reconstruction technique (SART), the maximum-likelihood (ML) and the total-variation regularized least-square reconstruction method (TVLS). Quality of the images reconstructed from these three methods is assessed through task-based performance. Two tasks are considered in this work: lesion detection and shape discrimination. Area under the ROC curve (AUC) is used as the figure-of-merit. Our simulation results indicate that TVLS and SART perform very similarly and better than the ML in terms of lesion detectability, while the ML performs better than the other two in terms of shape discrimination ability.