Prediction of facial profiles following orthodontic treatment in adult Class III patients using deep neural networks with feature fusion.

OBJECTIVE

This study proposed a cascaded framework based on deep neural networks that integrate cephalogram and landmark data to predict soft tissue profiles following orthodontic treatment in adult Class III patients.

METHODS

Paired pretreatment and posttreatment lateral cephalograms from 334 adult Class III patients were annotated and superimposed to calculate landmark coordinates. A deep neural network consisting of ResNet-50 and fully connected layers was developed and separately trained to predict skeletal changes from pretreatment cephalogram and incisor movement, and to predict soft tissue changes incorporating the skeletal changes as an additional predictor. Model performance was evaluated under two modes: using actual skeletal changes (Test Mode 1) or predicted skeletal changes (Test Mode 2) as the corresponding predictors.

RESULTS

The mean Euclidean errors for all landmarks were below 2 mm except Me' (2.20 ± 1.28 mm). In the upper lip region, the overall Euclidean error showed no statistical difference between Test Modes 1 and 2 (1.33 ± 0.80 mm vs. 1.37 ± 0.77 mm, P = 0.503). In the lower lip-chin region, Test Mode 1 showed a statistically lower error than Test Mode 2 (1.54 ± 1.17 mm vs. 1.83 ± 1.12 mm, P = 0.019).

CONCLUSIONS

The proposed network and workflow effectively predict posttreatment soft tissue profiles in adult Class III camouflage patients. Targeted optimization of either skeletal change prediction or hard-to-soft response prediction in specific anatomical regions can be a viable strategy for improving soft tissue prediction accuracy.

CLINICAL SIGNIFICANCE: This framework could serve as an innovative approach to improving the accuracy of facial profile prediction, thereby better assisting orthodontists as well as oral and maxillofacial surgeons in treatment planning and patient communication.