Two-level model averaging techniques in Drosophila brain imaging

For molecular brain mapping research, a reference template is necessary for integration of experimental data from different laboratories. However, the concept of the conventional brain atlas cannot provide such a reference template that fulfills our requirements. We propose a specific algorithm for model averaging to construct a 3D reference template. It extracts the geometric information of all data sets by reconstructing the individual wireframe models, and extracts their skeletons for determining the average skeleton. Large-scale averaged models, pseudoaverage models, can be obtained by warping the individual models toward the average skeleton. The final average model can be yielded after a small-scale model averaging procedure that determines the geometric median of pseudoaverage models. In contrast with other algorithms, our algorithm allows more severe variations because of its two-level hierarchy.