Uncertainty-Guided Joint Affine and Deformable Registration for Dental CBCT Images

• 1、School of Computer Science ＆ Technology，Beijing Jiaotong University, Beijing 100044

Abstract：Image registration is a key technique for achieving spatial alignment between images through geometric transformations and plays a crucial role in medical image analysis. Although deep learning-based methods have made significant progress, research on Cone Beam Computed Tomography (CBCT) image registration for orthodontic treatment remains relatively limited. The complex morphological changes induced by orthodontic treatment pose specific challenges for accurate registration of such images. To address these challenges, this paper proposes a novel Uncertainty-Guided Registration Network (UGR-Net). The main contributions are as follows: (1) introducing pseudo-segmentation labels as guidance to achieve coarse-to-fine affine and deformable registration within a unified network; (2) designing an uncertainty-guided weighting module to constrain the initial affine transformation by measuring feature similarity between image pairs, thereby enhancing registration robustness; (3) constructing the first dedicated CBCT dataset for orthodontic registration, named DentalReg, to provide a benchmark for objective evaluation. Experimental results on DentalReg show that the proposed method outperforms existing algorithms in affine registration and achieves competitive performance in deformable registration compared to mainstream methods.

Keywords： Deep Learning Image Registration Weakly-supervised Registration