The use of high resolution peripheral quantitative computed tomography (HR-pQCT) and in vivo micro-CT for studies of bone disease and treatment has become increasingly common, and with these methods comes large quantities of data requiring analysis. A simple, robust, and fully-automated segmentation algorithm is presented that efficiently segments bone regions. The dual threshold technique refers to two required threshold inputs that are used to extract the periosteal and endosteal surfaces of the cortex. The proposed method was tested against the gold standard, semi-automated hand contouring, using 45 datasets:​ mouse, rat, human, and cadaver data from the tibia or radius with nominal isotropic resolutions of 10–82 μm.

The performance of the proposed method to segment cortical and trabecular compartments was evaluated qualitatively from visualizations and quantitatively based on morphological measurements. Visual inspection confirmed successful segmentation of all datasets using the new method, with qualitatively better results when applied to the human and cadaver data compared to the gold standard. The dual threshold algorithm was able to extract thin and porous cortices, whereas some clipping and perforations occurred for the gold standard.

Morphological parameters measured for segmentation by the proposed method versus the gold standard agreed (95% confidence) for Tb.Th, Tb.Sp, and Tb.N, but not Ct.Th and BV/TV for the human and cadaver datasets. Nonetheless, correlations ranged from 0.95 to 1.00 for all morphological parameters except the cadaver Ct.Th because systematic errors were present. Poor agreement for Ct.Th and BV/TV was due to qualitatively incorrect segmentation by the gold standard when the cortex was thin compared to trabeculae, or operator bias during hand contouring. Since Tb.Th, Tb.Sp, and Tb.N were insensitive to segmentation method, despite operator bias, they are robust parameters for inter-site comparisons.

The dual threshold method offers a robust and fully-automated alternative to the gold standard that can efficiently segment bone regions with accurate and repeatable results. The algorithm can be easily implemented since it uses simple image analysis tools. Two input thresholds allow adjustment of the masked output, and are easily determined by trial and error. Using the same input thresholds for similar datasets assures maximal consistency while alleviating time consuming semi-automated contouring.