Speaker
Description
We propose a 3D surface extraction benchmarking tool to evaluate the performance of 3D model extraction methods from 2D discrete label maps. We have used one geometrical and one anatomical model as references. The anatomical model is selected from the BigBrain hypothalamus atlas (Jones et al. in prep).
3D triangular meshes are used as ground truth 3D surfaces. These contain various geometric features, such as concave and convex curvatures, textures, smooth surfaces, etc. 2D label maps of these reference models are extracted using Atelier3D software (A3D; Borgeat et al., 2007) and 3D surfaces are reconstructed from these using a surface extraction algorithm applying two methods: method 1 uses 2D label maps of every 20 sections, and method 2 includes every 2 sections from the BigBrain sectioning planes to reconstruct the shape of the structure. The reconstructed model is then evaluated by comparing to the original (reference) 3D model, considering features such as Hausdorff distance, RMS (root mean square), volume, surface area, number of vertices, and number of polygons.
Increasing the resolution (by including more sections) improves the performance of 3D surface extraction. This tool allows for interoperability of common visualization and annotation tools used around BigBrain, and provides the ability to compare the performance of different algorithms for 3D model, surface, or mesh creation processes from 3D volumes/2D discrete label maps at different resolutions. Such a benchmarking approach facilitates collaboration, helps improve the accuracy and scalability of 3D surface extraction, and promotes reproducible research.
