Dual-energy computed tomography (CT) offers the potential to recognize material properties by decomposing sinograms into Compton and photoelectric bases and subsequently reconstructing the Compton and photoelectric images. However, the presence of high density materials such as metal can distort the reconstructed images, leading to inaccurate material characterization. In this paper, we present a reconstruction technique to reduce noise and metal artifacts in dual-energy CT images by exploiting (1) statistical correlation between measurements and decomposed sinograms, (2) intra-image correlation between decomposed images, and (3) inter-image sparsity. The algorithm is based on minimizing weighted least squares with edge-preserving total variation regularization and is solved using split-Bregman iterative techniques. Using experimental data acquired from a commercial scanner, we demonstrate that the proposed algorithm significantly reduces noise and metal artifacts compared to the baseline approaches of filtered back projection and competing iterative reconstructions algorithms.