Novel radiographic stem version predictor from anterior-posterior radiographs.

Abstract

Implant orientation has been linked to postoperative outcomes and is frequently assessed on radiographs. However, the projection of the three-dimensional joint and implant structure to a two-dimensional radiograph complicates its assessment. The main objective of this study was to demonstrate a novel method for evaluating radiographic stem version, in a manner robust to multiaxial rotations, particularly AP tilt and flexion. Radiographic features where synthesised using a computational stem geometry and radiographic simulation, building in clinical error sources. Features trained a Gaussian process regression predictor of radiographic stem version. The impact of AP tilt on the accuracy of the Weber technique was then evaluated and the feasibility of AP tilt assessment from the same radiograph investigated. Radiographic stem version prediction accuracy was evaluated on in vitro radiographs with R² rising from 0.85 (P < 0.01) using the Weber technique to 0.98 (P < 0.01) using the trained model. Similar results were observed in a larger in silico dataset with R² rising from 0.89 (P < 0.01) to 0.98 (P < 0.01). Tilt was shown to reduce the accuracy of the Weber technique. Projectional symmetry was then demonstrated about the femoral implant with AP tilt, elucidating ambiguity when assessing tilt on an AP radiograph. The novel feature-based method is a reliable measure of radiographic stem version that is robust to variation on multiaxial orientation, allowing assessment of changing rotation in series of postoperative radiographs. However, a controlled radiograph is required to ensure this mirrors implanted stem version.