Comparative Study on Methods for Scaphoid Bone Model Completion from Sonography

Patient-specific bone models are required for surgical planning of computer-assisted percutaneous scaphoid fracture fixation. 3D sonography may present an alternative to computed tomography (CT) or magnetic resonance imaging (MRI) for the acquisition of those bone models, but it requires a completion process to derive full models from partial sonographic surfaces. To date, methods based on statistical shape models (SSMs) represent the state-of-the-art for this completion process. However, we have shown feasibility of a deep learning (DL)-based method for scaphoid bone model completion in a previous study. In this study, we compare our DL-based approach against three SSM-based completion methods: Active shape models (ASM), least squares optimization (LSO) and general-purpose optimization (GPO). 85 scaphoid bone models were used for training the DL-based AdaPoinTr as well as for building the SSM. All completion methods were evaluated on 20 additional test models, with partial input point clouds generated using a subsampling algorithm that mimics 3D sonography. Evaluation in terms of symmetric surface distance between completed mesh and corresponding ground truth mesh showed 1.1 mm for ASM, 0.7 mm for GPO, 0.5 mm for LSO and 0.3 mm for AdaPoinTr. The assessment of suitability for screw planning showed 12 protruding screws for ASM, 4 protruding screws for GPO and LSO each, and no protrusion for AdaPoinTr. Also, AdaPoinTr was found to be at least one order of magnitude faster than all other methods. Nevertheless, SSM-based completion methods may be better suited for smaller datasets and if the generation of plausible shapes is to be ensured.