Statistical iterative reconstruction (SIR) algorithms have shown advantages over the conventional filtered back-projection method for low-dose computed tomography (CT) reconstruction. For the SIR algorithms, the regularization term plays a critical role on determining the performance. One commonly used regularization is the quadratic-form Gaussian Markov random field (MRF), which penalizes differences among neighboring pixels in a small fixed window without considering discontinuities in images, thus may lead to over smoothing of edges or fine structures. In this work, we presented a quadratic-form MRF-based regularization with varying window size determined by the object scale, which is a descriptor of the image uniformity. For a uniform region (object scale is large), a larger MRF window is adopted because the coupling between the central pixel and its neighbors is strong; while for the interface region (object scale is small), a smaller MRF window is employed since the coupling is weak. The presented regularization term is incorporated into the penalized weighted least-squares (PWLS) iterative reconstruction scheme to improve low-dose CT reconstruction. Simulation results with a Shepp-Logan phantom revealed the presented regularization term is superior to the conventional Gaussian MRF in terms of noise suppression and edge preservation.