Teeth segmentation via semi-supervised learning

Efficient dental segmentation from volume data provides important assistance for orthodontic surgery and treatment. However, this task exits several major challenges due to the special dental anatomy and topological changes. This paper presents a robust interactive dental segmentation method, which treats this problem as a semi-supervised learning task. An initial classification of 3D mean shift is performed to partition the volume data into homogeneous blocks to guide the subsequent learning. It is easy to implement because only some simple operations are needed. It is accurate because a more general linear or nonlinear model can be learned by virtue of semi-supervised learning. Experimental results demonstrate the performance of the proposed scheme in extracting dental contours from complex background.