Performance Evaluation of Deep Learning Models for Practical 3D Tooth Scan Data Segmentation: Application of MeshSegNet and DilatedToothSegNet Models

This study aims to improve the clinical applicability of 3D segmentation models for tooth numbering by collecting and labeling 910 3D tooth scans, including prosthetic (crown, implant, inlay) and localized scan data, into 15 classes (gingiva and individual tooth numbers). Previous studies often relied on full oral structure data, failing to capture the irregularities of real clinical scenarios. To address this limitation, this research compares the performance of two deep learning models, MeshSegNet and DilatedToothSegNet, for tooth-numbered 3D segmentation. The models were trained with mesh cell resolutions of 10,000 and 16,000 to assess accuracy at different levels. Evaluation metrics, including Dice Similarity Coefficient (DSC) and mean Intersection over Union (mIoU), were used to measure performance. Results showed that the DilatedToothSegNet model with 16,000 cells outperformed MeshSegNet, achieving a 2.2-3.5% improvement in both DSC and mIoU metrics for maxillary and mandibular data. Additionally, it demonstrated consistent and reliable performance on localized scans, including prosthetic data.