Conventional cluster analysis of gene expression is often limited in its ability to incorporate cellular level heterogeneity that exists in the brain. We generate in situ hybridized gene expression cellular resolution maps (a set of multiple 2D images for each gene) of the mouse brain. Using a digital mouse brain atlas and advanced image analysis methods, gene expression profiles for each brain structure is calculated. We present a method to identify brain structure clusters with similar expression for a given gene using multivariate model-based clustering. In this study a family of Gaussian mixture models is used. Variation in the model is derived from parameterizing the covariance matrix by the shape, volume and orientation. Using expectation maximization and Bayesian information criterion both optimal model parameters and the number of clusters are determined. The results facilitate effective identification of brain structures with biologically interpretable expression profiles in a fully automated manner.