Improved two-level model averaging techniques in Drosophila brain modeling

Two-level model averaging techniques have been proposed to construct the 3D reference template for the Drosophila brain. The surface-based reference template is suitable for integration of experimental data from different laboratories. The 3D distance transform is the most memory and time consuming part in the model averaging algorithm. With the improvement of microscopic scanning technology, images of higher resolution can be acquired. Thus, the memories required for 3D distance transform become critical. In this paper, improved two-level model averaging techniques are proposed with three improvements. A two-scale distance map creation algorithm is introduced to reduce the memory cost in the distance transform. The computational time is reduced by a reduction of computation points in the distance map creation. The third improvement is an outlier rejection module to improve the robustness of the resulting average model.