Regularization parameter selection is crucial in CT iterative reconstruction because it is a balance between fidelity and penalty while there has not been so far a metric to judge if an optimal selection is made which results in the best reconstructed image quality in terms of a well balance between the apparent noise and the image fine structure. In this paper, we proposed a metric for selecting the optimal regularization parameter based on the property of natural image statistics. By using LoG operator and the pairwise products of neighboring LoG signals to extract the statistic features of the reconstructed image to account for the image quality, this proposed method evaluated all selected regularization parameters by calculating the variance of extracted statistic features and picked up the optimal regularization parameter with maximum curve of second order derivation of the calculated variance curve. Numerical and experimental results validated the efficiency of the proposed metric in terms of that the selected regularization parameter is accordance with the best visual observation. Besides, the proposed metric has low complexity of computation and only depends on features, which can be used in multiple situations.