Cavity Segmentation in X-ray Microscopy Scans of Mouse Tibiae

Osteoporosis is a chronic disease that causes lower bone density and makes bones fragile. This severely impairs patient life qualities and increases the burden on the social and health care system. X-ray microscopy (XRM) allows tracking of osteoporosis-related changes at a microstructural level in the bone, entailing the characterization of osteocyte lacunae and blood vessel canals. Unfortunately, no segmentation methods for micro-structures in XRM images have yet been established. In this work, we compare the performance of a traditional thresholding-based method with three deep learning networks including 2D and 3D models in both binary and multi-class segmentation. We further propose a clustering method to automatically distinguish blood vessels from lacunae for the binary methods. The performance is evaluated with Dice score (F1 score). The thresholding-based method reaches a mean Dice score of 0.729, which the deep learning models improve by 0.129 - 0.168.