Statistical content-adapted sampling (SCAS) for 3D Computed Tomography

In this paper, a framework to create a statistical content-adapted sampling (SCAS) for 3D X-ray Computed Tomography (CT) is introduced. SCAS aims at providing an accurate but light reconstruction volume. Based on decision theory, the 3D reconstruction space is sampled from the raw projection data in three steps to directly fit the sample. To do so, the structural information is first extracted from the projections by edge detection. This information is then merged in the reconstruction space, providing a pointcloud which accurately delineates the 3D interfaces of the specimen. From this pointcloud, a 3D mesh, closely fitting the shape of the studied object, is finally built via constrained Delaunay tetrahedralization. To assess the potential of the proposed SCAS for CT imaging, an iterative reconstruction was performed by classical Ordered Subset Simultaneous Algebraic Reconstruction Technique (OS-SART) - with fitting projection operator. The SCAS was evaluated on both numerical and experimental data. Results show that the use of statistical testing enabled the design of a robust, automated and fast method to build accurate pointclouds from a limited number of projections. The 3D meshes generated from these pointclouds are composed of few cells when compared to the regular voxel representation, leading to a downsize in computational cost and achieving up to 90% of memory footprint reduction. Simulations showed that performed reconstruction on such meshes provide accurate description of the object due to the finer sampling at interfaces.