Dental midline localization: development and validation of an automated deep learning pipeline

To develop and validate an automated pipeline for maxillary dental midline localization on posed smile photographs and compare the performance of four convolutional neural network (CNN) backbones.

A total of 221 standardized frontal posed smile photographs were collected. The maxillary dental midline was annotated by two orthodontists across two sessions, and mean coordinates served as the orthodontist-derived reference standard. An automated pipeline including face detection, interpupillary-line rotational normalization, mouth region extraction, and CNN-based keypoint regression was developed. Four CNN models (EfficientNet-B0, ResNet-18, MobileNetV3-Small, and HRNet-W18) were trained and compared using 5-fold stratified cross-validation. Deviation was expressed as the angle between the predicted dental midline and the interpupillary-derived photographic reference line (T-line).

HRNet-W18 achieved the lowest mean angular error (0.63° ± 0.63°) and the highest ≤1° error-threshold accuracy (81.0%). All 4 models exceeded 90% accuracy at the ≤2° threshold. No significant effect of tooth visibility grade on angular error was observed. HRNet-W18 also showed the highest agreement with the reference standard [intraclass correlation coefficient (ICC) = 0.721] and minimal Bland-Altman bias (+0.03°). Post hoc pairwise comparisons showed that HRNet-W18 outperformed all other models across all error metrics.

The sample was limited to standardized posed smile photographs from a single center, and generalizability to diverse populations and less controlled conditions remains to be established.

The automated pipeline showed high accuracy for dental midline localization on posed smile photographs relative to an orthodontist-derived reference standard. Further validation on larger and more diverse datasets is needed before clinical application.