Oral Cone Beam Computed Tomography Images Segmentation Based On Multi-view Fusion

Abstract

In computer-aided orthodontic treatment, it is necessary to establish a 3D model of the teeth and alveolar bone. Segmenting the teeth and alveolar bone from cone-beam computed tomography (CBCT) images is the primary step of reconstructing the model. However, previous studies mainly focused on the segmentation and reconstruction of teeth, and the research on alveolar bone segmentation is very rare. Different viewpoints of oral CBCT images provide complementary information. This paper proposes a multi-view information fusion method to realize the segmentation of teeth and alveolar bone in oral CBCT images. Firstly, for each of the three axes of the 3D data, an individual 2D deep network is trained to achieve coarse segmentation results. Secondly, in order to make full use of the information obtained from different viewpoints of oral CBCT images, a convolution network is used for multi-view fusion. To better fuse the information of different viewpoints, the channel-spatial attention mechanism is applied. Final segmentation is obtained by stacking 2D slices predictions, and the contextual information is added in a post-processing manner through fusing three-viewpoints results. According to the experimental results, our proposed method achieves significant improvements over the baseline.