Finding Dense Supervoxel Correspondence of Cone-Beam Computed Tomography Images

Dense correspondence establishment of cone-beam computed tomography (CBCT) images is a crucial step for attribute transfers and morphological variation assessments in clinical orthodontics. However, the registration by the traditional large-scale nonlinear optimization is time-consuming for the craniofacial CBCT images. The supervised random forest is known for its fast online performance, thought the limited training data impair the generalization capacity. In this paper, we propose an unsupervised random-forest-based approach for the supervoxel-wise correspondence of CBCT images. In particular, we present a theoretical complexity analysis with a data-dependent learning guarantee for the clustering hypotheses of the unsupervised random forest. A novel tree-pruning algorithm is proposed to refine the forest by removing the local trivial and inconsistent leaf nodes, where the learning bound serves as guidance for an optimal selection of tree structures. The proposed method has been tested on the label propagation of clinically-captured CBCT images. Experiments demonstrate the proposed method yields performance improvements over variants of both supervised and unsupervised random-forest-based methods.