AbstractsComputer Science

X-ray Computed Tomography (CT) is an effective nondestructive technology widely used for medical diagnosis and security. In CT, three-dimensional images of the interior of an object are generated based on its X-ray attenuation. Conventional CT is performed with a single energy spectrum and materials can only be differentiated based on an averaged measure of the attenuation. Multi-Energy CT (MECT) methods have been developed to provide more information about the chemical composition of the scanned material using multiple energy-selective measurements of the attenuation. Existing literature on MECT is mostly focused on differentiation between body tissues and other medical applications. The problems in security are more challenging due to the larger range of materials and threats which may be found. Objects may appear in high clutter and in different forms of concealment. Thus, the information extracted by the medical domain methods may not be optimal for detection of explosives and improved performance is desired. In this dissertation, learning and adaptive model-based methods are developed to address the challenges of multi-energy material discrimination for security. First, the fundamental information contained in the X-ray attenuation versus energy curves of materials is studied. For this purpose, a database of these curves for a set of explosive and non-explosive compounds was created. The dimensionality and span of the curves is estimated and their space is shown to be larger than two-dimensional, contrary to what is typically assumed. In addition, optimized feature selection methods are developed and applied to the curves and it is demonstrated that detection performance may be improved by using more than two features and when using features different than the standard photoelectric and Compton coefficients. Second, several MECT reconstruction methods are studied and compared. This includes a new structure-preserving inversion technique which can mitigate metal artifacts and provide precise object localization in the estimated parameter images. Finally, a learning-based MECT framework for joint material classification and segmentation is developed, which can produce accurate material labels in the presence of metal and clutter. The methods are tested on simulated and real multi-energy data and it is shown that they outperform previously published MECT techniques.