We have developed a two-stage Gauss-Newton reconstruction process with an automatic procedure for determining the regularization parameter. The combination is utilized by our microwave imaging system and has facilitated recovery of quantitatively improved images. The first stage employs a Levenberg-Marquardt regularization along with a spatial filtering technique for a few iterations to produce an intermediate image. In effect, the first set of iterative image reconstruction steps synthesizes a priori information from the measurement data versus actually requiring physical prior information on the interrogated object. Because of the interaction of the Levenberg-Marquardt regularization and spatial filtering at each iteration, the intermediate image produced from the first reconstruction stage represents an improvement in terms of the least squared error over the initial uniform guess; however, it has not completely converged in a least squared sense. The second stage involves using this distribution as a priori information in an iteratively regularized Gauss-Newton reconstruction with a weighted Euclidean distance penalty term. The penalized term restricts the final image to a vicinity (determined by the scale of the weighting parameter) about the intermediate image while allowing more flexibility in extracting internal object structures. The second stage makes use of an empirical Bayesian/random effects model that enables an optimal determination of the weighting parameter of the penalized term. The new approach demonstrates quantifiably improved images in simulation, phantom and in vivo experiments with particularly striking improvements with respect to the recovery of heterogeneities internal to large, high contrast scatterers such as encountered when imaging the human breast in a water-coupled configuration.