LEARNED MIXED MATERIAL MODELS FOR EFFICIENT CLUSTERING BASED DUAL-ENERGY CT IMAGE DECOMPOSITION

Penalized weight-least squares (PWLS) with basis material priors is a promising way to achieve high quality material decompositions for Dual-energy CT (DECT). This paper proposes a new method dubbed DECT-MULTRA for image domain DECT material decomposition that combines conventional PWLS estimation with regular-ization based on a mixed union of learned transforms (MULTRA) model. Our approach pre-learns from training data a common union of unitary transforms for all the basis materials’ patches, as well as a cross-material union of unitary transforms that captures relationships between the different basis material images. The proposed DECT-MULTRA algorithm efficiently obtains material decompositions by alternating between updating the material images and performing clustering of patches in the MULTRA model. Both these steps of the alternating algorithm have closed-form updates. Numerical experiments with the XCAT phantom show that the proposed method significantly improves image quality compared to the recent DECT-ST method that learns different sparsifying transforms for different basis materials and the DECT-EP approach that uses a non-adaptive edge-preserving hyperbola regularizer.