During shoulder arthroplasty the native functionality of the diseased shoulder joint is restored, this functionality is strongly dependent upon the native anatomy of the pre-diseased shoulder joint. Therefore, surgeons often use the healthy contralateral scapula to plan the surgery, however in bilateral diseases such as osteoarthritis this is not always feasible. Virtual reconstructions are then used to reconstruct the pre-diseased anatomy and plan surgery or subject-specific implants. In this project, we develop and validate a statistical shape modeling method to reconstruct the pre-diseased anatomy of eroded scapulae with the aim to investigate the existence of predisposing anatomy for certain shoulder conditions.

The training dataset for the statistical shape model consisted of 110 CT images from patients without observable scapulae pathologies as judged by an experienced shoulder surgeon. 3D scapulae models were constructed from the segmented images. An open-source non-rigid B-spline-based registration algorithm was used to obtain point-to-point correspondences between the models. The statistical shape model was then constructed from the dataset using principle component analysis. The cross-validation was performed similarly to the procedure described by Plessers et al. Virtual defects were created on each of the training set models, which closely resemble the morphology of glenoid defects according to the Wallace classification method. The statistical shape model was reconstructed using the leave-one-out method, so the corresponding training set model is no longer incorporated in the shape model. Scapula reconstruction was performed using a Monte Carlo Markov chain algorithm, random walk proposals included both shape and pose parameters, the closest fitting proposal was selected for the virtual reconstruction. Automatic 3D measurements were performed on both the training and reconstructed 3D models, including glenoid version, critical shoulder angle, glenoid offset and glenoid center position.

The root-mean-square error between the measurements of the training data and reconstructed models was calculated for the different severities of glenoid defects. For the least severe defect, the mean error on the inclination, version and critical shoulder angle (°) was 2.22 (± 1.60 SD), 2.59 (± 1.86 SD) and 1.92 (± 1.44 SD) respectively. The reconstructed models predicted the native glenoid offset and centre position (mm) an accuracy of 0.87 (± 0.96 SD) and 0.88 (± 0.57 SD) respectively. The overall reconstruction error was 0.71 mm for the reconstructed part. For larger defects each error measurement increased significantly.

A virtual reconstruction methodology was developed which can predict glenoid parameters with high accuracy. This tool will be used in the planning of shoulder surgeries and investigation of predisposing scapular morphologies.

Pre-operative 3D glenoid planning improves component placement in terms of version, inclination, offset and orientation. Version and inclination measurements require the position of the inferior angle. As a consequence, current planning tools require a 3D model of the full scapula to accurately determine the glenoid parameters. Statistical shape models (SSMs) can be used to reconstruct the missing anatomy of bones. Therefore, the objective of this study is to develop and validate an SSM for the reconstruction of the inferior scapula, hereby reducing the irradiation exposure for patients.

The training dataset for the statistical shape consisted of 110 CT images from patients without observable scapulae pathologies as judged by an experienced shoulder surgeon. 3D scapulae models were constructed from the segmented images. An open-source non-rigid B-spline-based registration algorithm was used to obtain point-to-point correspondences between the models. A statistical shape model was then constructed from the dataset using principal component analysis. Leave-one-out cross-validation was performed to evaluate the accuracy of the predicted glenoid parameters from virtual partial scans. Five types of virtual partial scans were created on each of the training set models, where an increasing amount of scapular body was removed to mimic a partial CT scan. The statistical shape model was reconstructed using the leave-one-out method, so the corresponding training set model is no longer incorporated in the shape model. Reconstruction was performed using a Monte Carlo Markov chain algorithm, random walk proposals included both shape and pose parameters, the closest fitting proposal was selected for the virtual reconstruction. Automatic 3D measurements were performed on both the training and reconstructed 3D models, including glenoid version, inclination, glenoid centre point position and glenoid offset.

In terms of inclination and version we found a mean absolute difference between the complete model and the different virtual partial scan models of 0.5° (SD 0.4°). The maximum difference between models was 3° for inclination and 2° for version. For offset and centre point position the mean absolute difference was 0 mm with an absolute maximum of 1 mm.

The magnitude of the mean and maximum differences for all anatomic measurements between the partial scan and complete models is smaller than the current surgical accuracy. Considering these findings, we believe a SSM based reconstruction technique can be used to accurately reconstruct the glenoid parameters from partial CT scans.

Summary Statement

In this study, excellent positioning of custom-made glenoid components was achieved using patient-specific guides. Achieving the preoperatively planned orientation of the component improved significantly and more screws were located inside the scapular bone compared to implantations without such guide.