Deep reinforcement learning for efficient registration between intraoral-scan meshes and CT images

Abstract

Registration between computed tomography (CT) images and intraoral-scan (IOS) meshes facilitates dental procedure planning. However, the spatial complexity of 3D-space computations presents a significant challenge, necessitating the reduction of computational cost through efficient sampling while maintaining robustness via global approximation without segmentation. Herein, we introduce an efficient and robust method for registering CT images and IOS meshes, eliminating the need for segmentation. We utilized an effective sampling technique to identify key vertices in IOS meshes by calculating the negative curvatures between adjacent faces. The significant vertices are transformed into a novel graph representation, serving as the input state for the graph convolution-based backbone network within a deep reinforcement learning (DRL) framework. This framework approximates an optimal solution through sequential decision-making, selecting the best among 12 actions by considering translation and rotation to accurately locate the 3D mesh at arbitrary positions and angles on maxillary or mandibular teeth in CT images. The proposed method was evaluated against conventional and deep learning-based methods, demonstrating mean absolute errors of 1.955 ± 1.310 and 1.399 ± 0.644 mm for maxillary and mandibular teeth, respectively. Additionally, it required only 0.48 M floating-point operations for the calculations, making it more efficient than existing methods.