We consider the problem of dendritic spine detection from X-ray micro-tomographic volumes that allow huge volume of tissue visualization. To compensate for the noise in data that induces false positives in the spine detection process, we first segment the dendrites. This segmentation is obtained by computing the medial axis and approximating the results by segments obtained with a 3D Hough transform. Dendrites are then reconstructed and a spine mask is obtained using the typical diameter of dendrites and distance between spine and dendrites. A point process is then optimized on this mask, thus providing the spine detection.