Increasing use of patient-specific surgical procedures in orthopaedics means that patient-specific anatomical coordinate systems (ACSs) need to be determined. For knee bones, automatic algorithms constructing ACSs exist and are assumed to be more reliable than manual methods, although both approaches are based on non-unique numerical reconstructions of true bone geometries. Furthermore, determining the best algorithms is difficult, as algorithms are evaluated on different datasets. Thus, in this study, we developed 3 algorithms, each with 3 variants, and compared them with 5 from the literature on a dataset comprising 24 lower-limb CT-scans. To evaluate algorithms’ sensitivity to the operator-dependent reconstruction procedure, the tibia, patella and femur of each CT-scan were each reconstructed once by three different operators.
Our algorithms use principal inertia axis (PIA), cross-sectional area, surface normal orientations and curvature data to identify the bone region underneath articular surfaces (ASs). Then geometric primitives are fitted to ASs, and the ACSs are constructed from the geometric primitive points and/or axes. For each bone type, the algorithm displaying the least inter-operator variability is identified. The best femur algorithm fits a cylinder to posterior condyle ASs and a sphere to the femoral head, average axis deviations: 0.12°, position differences: 0.20 mm. The best patella algorithm identifies the AS PIAs, average axis deviations: 0.91°, position differences: 0.19 mm. The best tibia algorithm finds the ankle AS center and the 1st PIA of a layer around a plane fitted to condyle ASs, average axis deviations: 0.38°, position differences: 0.27 mm.