Tooth segmentation and dental crowding diagnosis using two-stage dual-dilated graph convolution.

Abstract

Purpose: Tooth segmentation and diagnosis of dental crowding severity on 3D intraoral scan models are key processes for computer-aided analysis of orthodontic models. Conventional methods are time-consuming, inefficient, and subjective, necessitating more efficient and intelligent approaches. Therefore, we propose a two-stage intelligent workflow.

Methods: In Stage 1, tooth segmentation is performed using an innovative dual-dilated graph convolutional network 1 (DDGCNet1). In Stage 2, Stage 1's output is converted to a point cloud, then processed by DDGCNet2 and post-processing to generate arch length discrepancy (ALD, an indicator of dental crowding). The encoding layers of the proposed networks embed a novel dual-dilated EdgeConv module, effectively learning from local features and long-range contextual information of adjacent teeth.

Results: Experimental comparative analysis demonstrates that the proposed network achieves outstanding segmentation performance and accurate dental crowding diagnosis. In ALD measurement, it attains a mean absolute error (MAE) of 1.553 mm for the maxilla and 1.434 mm for the mandible.

Conclusion: This study can assist orthodontists in diagnosis and treatment, alleviate their workload, and expedite the development of reliable orthodontic treatment plans, thereby meeting the demands of computer-aided orthodontic diagnosis.