Automatic Detection of Cortical Bone’s Haversian Osteonal Boundaries

This work aims to automatically detect cement lines in decalcified cortical bone sections stained with H&E. Employed is a methodology developed previously by the authors and proven to successfully count and disambiguate the micro-architectural features (namely Haversian canals, canaliculi, and osteocyte lacunae) present in the secondary osteons/Haversian system (osteon) of cortical bone. This methodology combines methods typically considered separately, namely pulse coupled neural networks (PCNN), particle swarm optimization (PSO), and adaptive threshold (AT). In lieu of human bone, slides (at 20× magnification) from bovid cortical bone are used in this study as proxy of human bone. Having been characterized, features with same orientation are used to detect the cement line viewed as the next coaxial layer adjacent to the outermost lamella of the osteon. Employed for this purpose are three attributes for each and every micro-sized feature identified in the osteon lamellar system: (1) orientation, (2) size (ellipse perimeter) and (3) Euler number (a topological measure). From a training image, automated parameters for the PCNN network are obtained by forming fitness functions extracted from these attributes. It is found that a 3-way combination of these features attributes yields good representations of the overall osteon boundary (cement line). Near-unity values of classical metrics of quality (precision, sensitivity, specificity, accuracy, and dice) suggest that the segments obtained automatically by the optimized artificial intelligent methodology are of high fidelity as compared with manual tracing. For bench marking, cement lines segmented by k-means did not fare as well. An analysis based on the modified Hausdorff distance (MHD) of the segmented cement lines also testified to the quality of the detected cement lines vis-a-vis the k-means method.