REFERENCE AREAS SELECTION AFFECTS REGISTRATION OF AI-SEGMENTED MANDIBLES ACQUIRED WITH CBCT

Abstract

Introduction Precise registration of sequential 3D datasets is crucial for accurate dimensional analysis. Utilizing the Local Best-Fit (LBF) algorithm and stable Registration Reference Areas (RRAs) facilitates the accurate alignment of 3D surface models. Currently, Cone-beam Computed Tomography (CBCT) and Deep Learning (DL) algorithms are at the forefront for segmenting CBCT scans to monitor morphological changes in the residual alveolar ridge. This study compares the effectiveness of different RRAs in registration sequential 3D surface models of partially edentulous mandibles. Methods DL-assisted software segmented two sequential CBCTs (T0 and T1) from 10 patients, producing sequential 3D mandibular models. These models were aligned using three distinct RRAs: (i) WHOLE, encompassing the entire surface model; (ii) MND_BODY, covering the mandibular body while excluding the unstable alveolar ridge; and (iii) SPIN_FOR, incorporating stable RRAs (mental foramina and mental spine). An innovative method assessed registration accuracy by generating centroids from cross-sectional outlines of the mandibular nerve canals at the anterior third (A), medial third (B), and posterior third (C) of the posterior edentulous areas. The distance between centroids at T0 and T1 quantified registration accuracy. Results The MND_BODY group exhibited superior accuracy, whereas the SPIN_FOR group showed the least, with accuracy decreasing from A to C, suggesting rotational misalignments. ConclusionsWhen selecting RRAs, both stability and spatial distribution must be taken into account. For optimal alignment, sequential 3D surface models should use RRAs that are both stable and widely distributed.