Computer-Assisted Osteotomy Planning: Medical Imaging Techniques for Orthopaedic Surgery

Abstract

In this thesis, the ultimate goal has been to strive towards (automatic) computer assisted planning of reconstructive osteotomies. However, the quality of the output of such a workflow depends on the quality of all the intermediary steps, between the acquisition of a medical image, to the virtual reconstruction of the bone. In each chapter of this thesis, we have tried to overcome some of the challenges posed by these steps, and to evaluate the intermediate results. First, a method to improve the registration of CT to MRI scans for knee joint images was developed. Large deformations can arise due to differences in patient positioning between scanners, making registration complex. The study showed that non-rigid registration using deformable B-Splines could be improved by initializing the registration with a rigid registration of bones and estimating soft tissue deformation using a dual quaternion-based interpolation step. Second, an automatic and efficient deep learning-based method for semantic segmentation of six separate bones from lower extremity CT scans was proposed and validated. Existing deep learning segmentation algorithms were found to be slow and memory-intensive for this application. The optimized architecture of a cascaded U-net approach outperformed the state-of-the-art nn-Unet in memory efficiency, speed, and accuracy. Automated morphometric measurements on the hip joint were clinically validated, showing good to excellent correlation with measurements derived from manual segmentations. Third, the impact of using CT and MRI to segment bone and/or cartilage on pre-operative osteotomy planning of the forearm in adolescent patients was evaluated. An automated, deterministic planning method was developed to directly compare planning outcomes and avoid inter- or intraoperator variation. Segmentations of bone from both CT and MRI were used to assess the impact of different modalities on planning. Additionally, segmentations of bone and bone combined with cartilage from MR images were analyzed to study the influence of cartilage on planning outcomes. Results showed excellent correlation between realignment parameters across different segmentations, with small significant differences in the translational part of realignment. A positive correlation was found between cartilage amount and differences in planning on bone and bone with cartilage, suggesting the importance of considering cartilage impact on joint shape in young patients. However, further research is needed to explore MRI as a potential replacement for CT in osteotomy planning due to the small sample size and time-consuming manual segmentation. Lastly, a method to automatically quantify alignment parameters describing lower extremity morphology from 3D bone models was evaluated. A completely automatic pipeline was developed to extract these parameters using a CT scan as input. The alignment parameters were calculated using landmarks from previous studies and expanded upon to quantify all parameters needed for lower extremity realignment planning. This automatic method eliminated the need for manual operation and produced results that closely corresponded to intrarater variability. This advancement could pave the way for automatic realignment planning of lower extremities, addressing cases where a healthy contralateral example is unavailable for unilateral bone realignment planning.