Automatic segmentation of cortical bone microstructure: Application and analysis of three proximal femur sites.

Osteoporosis is the most common bone metabolic unbalance, leading to fragility fractures, which are known to be associated with structural changes in the bone. Cortical bone accounts for 80 % of the skeleton mass and undergoes remodeling throughout life, leading to changes in its thickness and microstructure. Although many studies quantified the different cortical bone structures using CT techniques (3D), they are often realised on a small number of samples. Therefore, the work presented here proposes a method to quantify cortical bone microstructure using 2D histology, shows its application on a set of 94 samples and compares to 3D methods. Fresh frozen human femur pairs from 47 donors aged between 57 and 96 years were obtained from the Medical University of Vienna. Bone samples were cut from 3 sites: proximal part of the diaphysis, inferior and superior segments of the neck. The samples were stained with toluidine blue and imaged under light microscopy. After manual segmentation of a few regions of interest by multiple operators, a convolutional neural network was trained in combination with a random forest for automatic segmentation. The segmentation analysis compares morphology and structure distribution of Haversian canals, osteocyte lacunae, and cement lines with literature, between anatomical sites, sex, left and right sides, and relation to ageing. Morphological analysis of the segmentation gives results similar to the literature. Comparison between male and female donors shows no significant differences. There is no significant difference between left and right femur on paired samples but significant differences are observed between anatomical locations. The structures' relative amounts do not present significant changes with age but only weak tendencies. Nevertheless, a strong correlation was observed between osteocyte lacunae density and bone areal fraction. This study presents a full process to stain and automatically segment digital cortical bone images. Its application to a large sample set of proximal femora provides strong statistics on the cortical bone structures morphology and distribution. Similarities observed between sides and sexes together with differences observed between sites could indicate that mechanical loading might be a main driver for bone microstructure. Additionally, the relationship between osteocyte lacunae density and bone areal fraction could suggest that bone porosity is regulated by osteocyte survival.