With the image processing techniques, nowadays it is possible to obtain 3D X-ray tomography which is used in medical diagnosis. Such systems require a reconstruction of an object - a human body in this case- in 3D from a set of its 2D projections. The reconstruction volume is usually discretized on a regular grid of isotropic voxels which implies an increase in their number to achieve good spatial resolution. In this work, we propose a process to obtain 3D mesh generation. method allowing to discretize the 3D reconstruction space in a relevant way directly from the structural information contained in the projection data. The idea is to obtain a representation adapted to the studied object. Here we have recourse to a tetrahedral mesh matching the structure of the object: the mesh density adapts according to the interfaces and homogeneous regions. To build such a mesh, the first step of the method consists in detecting the edges in the 2D projection data. The structural information thus obtained is then merged in the reconstruction space in order to construct a point cloud sampling the 3D interfaces of the imaged object.