Automated Multiclass Bone Segmentation Using Deep Learning: Implications for Templating in Radial Head Replacement.
Authors
Affiliations (6)
Affiliations (6)
- Department of Orthopedic Surgery, Mayo Clinic, Rochester, MN, USA; Clinica Universidad de los Andes, Department of Orthopedic Surgery, Santiago, Chile.
- Department of Orthopedic Surgery, Mayo Clinic, Rochester, MN, USA.
- Department of Orthopedic Surgery, Mayo Clinic, Rochester, MN, USA; Department of Orthopaedic Surgery, Showa University School of Medicine, Tokyo, Japan.
- Department of Radiology, Anatomic Modeling Unit, Mayo Clinic, Rochester, MN, USA.
- Department of Radiology, Anatomic Modeling Unit, Mayo Clinic, Rochester, MN, USA; Division of Neuroradiology, Department of Radiology, Biomedical and Scientific Visualization, Mayo Clinic, Rochester, MN, USA.
- Department of Orthopedic Surgery, Mayo Clinic, Rochester, MN, USA. Electronic address: [email protected].
Abstract
Preoperative three-dimensional (3D) templating can improve surgical accuracy in anatomic-press-fit radial head arthroplasty (RHA). However, current imaging segmentation methods used for templating are time-consuming and prone to variability. This study aimed to train and validate an nnU-Net deep learning model to automate multiclass bone segmentation for RHA templating. We hypothesized that the nnU-Net model would achieve high accuracy in segmenting the upper extremity bones thereby supporting 3D bone templating in RHA. A total of 93 upper extremity computed tomography (CT) scans met the eligibility criteria. Ground-truth segmentation was performed by a trained orthopedic surgeon and reviewed by a radiologist and an engineer to ensure accuracy. The nnU-Net model was trained and evaluated using the Dice Similarity Coefficient (DSC) and Hausdorff Distance to measure overlap and segmentation accuracy against manual segmentations. The 3D bone models derived from the nnU-Net model and manual segmentation were compared through Mean Surface Distance (MSD) and Root Mean Squared Error (RMSE) were determined to assess the surface variation between the bone models. The average time on segmenting each CT was compared. The nnU-Net achieved high segmentation accuracy with DSC values of 0.99 for the humerus, 0.98 for the ulna, and 0.96 and 0.95 for the cortical and non-cortical radii, respectively. The MSD remained below 0.2 mm for all bone classes. The mean RMSE values were consistent at 0.2 mm across all bones. Segmentation time averaged 3 min per scan compared to 78 min for manual segmentation, with consistent performance across gender, arm side, and CT slice thickness. This deep learning model provides a fast and reliable solution for multiclass bone segmentation and demonstrates high accuracy in segmenting cortical and non-cortical regions, which are essential for RHA templating. The accuracy was consistent with clinical needs and fits below the sizing intervals of commercially available prostheses. This supports its potential utility for 3D preoperative planning in RHA, despite its inability to capture cartilage. This approach demonstrates clinical feasibility for improving efficiency and precision in templating radial head replacement surgery.