Soft tissue prediction: conditional generative adversarial network versus Dolphin visual treatment objective in Asian adult class I extraction therapy

This study compared the accuracy and clinical applicability of a deep learning-based artificial intelligence (AI) model and the Dolphin visual treatment objective (VTO) to predict soft tissue changes after premolar extraction in patients with skeletal Class I malocclusion to provide empirical evidence for guiding prediction tool selection.

Thirty adult patients with skeletal Class I malocclusion who underwent extraction orthodontic treatment were enrolled. Pre-treatment (T1) and post-treatment (T2) lateral cephalometric images were collected. A conditional generative adversarial network (CGAN) model modified with a U-Net generator and PatchGAN discriminator was constructed to predict post-treatment soft tissue morphology. The AI model was pre-trained using cephalometric data from 511 patients. Tooth movement vectors from T1 to T2 served as model inputs. Data augmentation, early stopping, L2 regularization, and dropout layers were applied to mitigate overfitting. After training, 30 patients were included as an independent test set. Dolphin VTO served as the conventional comparison method. Soft tissue cephalometric analysis was conducted on AI-predicted, Dolphin VTO-predicted, and actual T2 images to measure landmark 2D Euclidean distances and soft tissue cephalometric indices. For qualitative validation, 15 random cases were evaluated by four calibrated raters (blinded to the prediction methods) using a 100-mm visual analog scale (VAS) to assess the similarity of predicted and actual facial profiles. Two-way multivariate analysis of variance (MANOVA) was used to compare VAS scores.

Intra- and inter-investigator reliabilities for landmark identification showed good consistency. For each case, the CGAN model achieved an inference time of approximately 2 s, whereas Dolphin VTO required approximately 5-10 min of manual operation. Compared with Dolphin VTO, the AI model produced significantly lower total mean landmark prediction error and had smaller prediction deviations from actual values for all soft tissue cephalometric indices. MANOVA revealed significant between-method differences in perceptual ratings. The AI model showed significantly higher VAS scores for profile, upper lip, and lower lip predictions, but without a significant difference in subnasal and chin region ratings. Rater's professional background had no significant interactive effect on the evaluation.

The CGAN-based AI model outperformed Dolphin VTO in predicting soft tissue changes and facial profile esthetics after extraction treatment of skeletal Class I malocclusion, with clinically acceptable prediction errors. This AI model may serve as an effective tool for predicting post-orthodontic treatment facial esthetics, facilitating clinical treatment planning, and doctor-patient communication.