Segmenting images of bones into cortical and trabecular regions is necessary to accurately analyse cortical and trabecular bone properties. A popular existing method created for long bones [1,2], uses a fixed kernel size closing cycle. However for vertebral bones with thin and highly curved cortices an inappropriate kernel size causes error due to incomplete trabeculae removal or excessive closing, eliminating highly curved cortical regions. We propose to reduce error by adjusting the fixed kernel size for vertebral bones, scaling the kernel size and modifying the method by adding a closing step before the opening cycle during the calculation of the endosteal surface.
5 human and kangaroo cervical vertebrae (C3-C7) were CT scanned at 35µm voxel size. The data was filtered, resampled at 70, 105 & 140µm voxel size and thresholded. Buie segmentation [1,2] was performed using kernel sizes from 1-20 voxels. The lowest kernel size with complete trabeculae removal was taken as the optimal kernel size. The cortical volume using the optimal kernel size is compared against a fixed kernel size, a scaled fixed value taken from the Buie study and the modified method.
The average/maximum optimal kernel size was 0.36/0.84mm and 0.91/1.54mm for the kangaroo and human samples respectively. The fixed method erroneously excluded 15%, 24% & 33% of cortical volume at 70, 105 & 140µm voxel size respectively, which was more than the scaled method at large voxel sizes to 16%, 22% & 27%. The modification to the scaled method further reduces error to 2%, 3% & 6%.
The optimal kernel size calculated for vertebral bones differs from the value determined for long bones in the previous study. Although some error can be avoided using a scaled kernel size and our closing cycle modification, to eliminate error the optimal kernel size must be found for each sample.