Automated alveolar bone and teeth 3D segmentation on the cone-beam computed tomographic images using the convolutional neural network

Abstract

Abstract Background Cone-beam computed tomography (CBCT) is widely used in dental diagnosis and implant planning. The accurate segmentation of teeth and alveolar bone from CBCT images is essential for precise diagnosis and treatment planning. In this study, we propose two distinct convolutional neural network models designed to effectively perform precise segmentation of alveolar bone and teeth Methods The first model is a multi-class segmentation convolutional neural network, termed 'Attentional Relation Unet++' (AR Unet++), specifically designed for slice-by-slice separation of the jawbone and tooth structures. The AR Unet + + enhances the original Unet + + model by incorporating a dual attention network and a local relation layer. This advancement allows the model to effectively integrate local features, as well as positional and channel attention. Complementing the AR Unet++, we introduce the 3D merged selective U-net (MS U-net), designed specifically for segmenting individual teeth within volumetric CT images. The MS U-net leverages a 2D/3D feature merge and selective kernel convolution to integrate across channels and diverse kernels. Results Our experiments utilized 9,330 axial slices and 973 individual tooth segments derived from forty anonymized CBCT volumes. The AR Unet + + attained average values of 0.9892, 0.0223, and 0.0016 for the dice similarity coefficient (DSC), relative volumetric overlap error (VOE), and relative volume difference (RVD), respectively. In the experiments conducted with the 3D MS-net, the performance indices observed were 0.9746 for the average DSC, 0.9508 for the average Jaccard coefficient (JC), 1.3263 mm for the average Hausdorff distance (HD), and 0.3180 mm for the average symmetric surface distance (ASSD). Conclusions We proposed two novel CNNs: the attentional relational U-net ++ (AR Unet ++) and the 3D MS U-net. The AR Unet + + was designed to accurately delineate jawbones and teeth from CBCT images in a slice-to-slice manner. Their experimental results underline the substantial potential of the proposed system as an advanced tool for enhancing clinical examinations.