[Evaluation of nasal prosthesis automatic construction integrating a flexible point cloud deformation matching algorithm and generative artificial intelligence].

Abstract

Objective: To develop a point cloud non-rigid deformation algorithm, local projection deformation match (LPD-Match), for optimizing the edge transition of the nasal prosthesis complementing the generative artificial intelligence-based facial mesh generation network (FMGen-Net) model previously developed by the research team to automate the construction of target reference data for the nasal prosthesis. Methods: The LPD-Match algorithm was developed based on the local projection deformation strategy. Three-dimensional facial scans from 20 subjects without significant facial deformities who first visited the Department of Prosthodontics, Peking University School and Hospital of Stomatology, between February 2023 and November 2023 were collected, and full nasal defect data were simulated on these scans. To evaluate the effectiveness of LPD-Match algorithm in edge transition optimization, a paired experiment was designed. For each model, two methods for constructing target reference data were compared. The control group utilized FMGen-Net alone, while the experimental group integrated FMGen-Net with the LPD-Match algorithm. The "edge fitness" metric was defined to evaluate the edge transition of the prosthesis by calculating the three-dimensional (3D) curve deviation and maximum curve deviation between the defect boundary curve and the prosthesis boundary curve. Additionally, the "morphological similarity" metric was introduced to assess the morphology restoration of the target reference data.Using the nasal region data from the original three-dimensional facial scans as the gold standard, the root mean square error of the three-dimensional morphological deviation and the maximum morphological deviation between the prosthesis design data and the gold standard were calculated for both groups. Results: Analysis of the "edge fitness" metric showed that the root mean square error of the 3D curve deviation in the control group averaged (0.62±0.23) mm, whereas in the experimental group, it was (0.16±0.07) mm. The maximum curve deviation in the control group averaged (1.57±0.64) mm, while the experimental group showed a reduced value of (0.52±0.30) mm. Statistical analysis showed significant differences between the two groups (t=11.33, P0.001; t=10.17, P0.001), with the experimental group showing superior edge transition. For the "morphological similarity" analysis, the root mean square error of the 3D morphology deviation was 1.30 (1.00, 1.83) mm in the control group and 1.23 (0.90, 1.63) mm in the experimental group. The maximum morphology deviation in the control group was 4.26 (2.96, 5.02) mm, while the experimental group demonstrated 3.85 (2.92, 4.56) mm. Statistical testing showed significant differences between the two groups (Z=-3.40, P0.001; Z=-2.09, P=0.037), indicating the superior morphological restoration effect of the experimental group. Conclusions: This study proposes a method using point cloud non-rigid deformation algorithms to optimize the edge transition of the nasal prosthesis and the preliminary verification demonstrates its feasibility and effectiveness. Compared with FMGen-Net model alone, the combined method integrating LPD-Match algorithm with FMGen-Net model improved the edge fitness of nasal prosthesis target reference data while maintaining favorable morphological reconstruction performance.The proposed method shows potential for personalized digital reconstruction of nasal defects in clinical prosthodontics.