Alveolar Bone and Mandibular Canal Segmentation on Cone Beam Computed Tomography Images Using U-Net

Abstract

A dental implant is a treatment to replace missing teeth. Determining the proper dimensions of dental implants is measured by observing the distance from the mandibular canal (MC) to the alveolar bone (AB). It is crucial to pay careful attention to the location of the MC when planning for a dental implant in the posterior mandible to avoid injury. Therefore, segmenting AB and MC in dental implant planning is essential. While research on MC segmentation using deep learning has been conducted extensively, there has yet to be much research on AB and MC segmentation simultaneously with deep learning. This study proposes using U-Net as a high-performance segmentation technique for multiclass and binary segmentation to segment AB and MC regions. In the output branch of the U-Net architecture, two scenarios are designed, the first is to perform AB and MC segmentation simultaneously, while the second is to perform AB and MC segmentation separately. The study used 563 2D grayscale Cone Beam Computed Tomography (CBCT) images from the coronal slice. The model is trained and tested using K-fold cross validation. The test results show that AB and MC segmentation simultaneously produces the mean intersection of union (IoU) value of 0.85. Meanwhile, AB and MC segmentation separately produced the mean IoU of 0.98 for AB segmentation and 0.81 for MC segmentation. The results of the satisfactory AB and MC segmentation are expected to assist determine implant dimensions in dental implant planning.