Single Bone Modeler: deep learning bone segmentation for cone-beam CT

The accurate segmentation and modeling of bones play a crucial role in diagnosis and surgical planning in orthopedics. Traditional methods face challenges in capturing the fine details and complex structures present in cone-beam computed tomography (CBCT) scans. This paper introduces a novel deep learning-based workflow to precisely segment bone in CBCT scans of complex areas such as extremities, the Single Bone Modeler (SBM). It involves three main steps: bone segmentation, separation and 3D modeling. To achieve highly accurate bone segmentation, a dedicated U-Net architecture is developed and compared to a SegNet. Furthermore, we compare two different training strategies axial training and multi-planar training, when dealing with CBCT data. The separation of bones is performed through a watershed algorithm, and the structure of interest is subsequently modeled in 3D. The efficacy of proposed deep learning approaches is assessed, and outcomes are compared to benchmark techniques using two metrics: Jaccard Index (JI) and Dice Coefficient (DC). Results demonstrate the superior performance in bone segmentation of the proposed U-Net trained with Multi-Planar training, achieving a JI of 0.941 ± 0.031 and a DC of 0.970 ± 0.015. The entire workflow is further evaluated for its capacity to isolate specific bone, showcasing significant improvement over benchmark methods. In conclusion, the SBM enhances the precision of bone segmentation in high-resolution CBCT scans. The results suggest the potential for reliable and efficient extremity bone segmentation, with implications for improved applications in orthopedics.