Accuracy of artificial intelligence-assisted soft tissue landmark identification in serial lateral cephalograms of Class III two-jaw surgery patients

Objective: To evaluate the accuracy of artificial intelligence (AI)-assisted soft tissue landmark identification (STLI) on serial lateral cephalograms (Lat-Cephs) of Class III patients treated with two-jaw orthognathic surgery across four different time-points. Methods: A convolutional neural network model was developed for STLI, trained and validated using 3,004 Lat-Cephs from 751 patients. The test set included 224 Lat-Cephs from 56 patients, divided into the genioplasty (n = 22) and non-genioplasty (n = 34) groups. The four time-points included initial (T0), pre-surgery (T1, brackets), post-surgery (T2, brackets, surgical plates, and screws [S-PS]), and debonding (T3, S-PS and fixed retainers). AI accuracy was compared with a human standard for 13 soft tissue landmarks. Mean radial errors (MREs), horizontal and vertical errors, and statistical differences were analyzed. Results: The total MRE across all time-points was 1.50 ± 0.48 mm, with 64.9% of values being less than 1.5 mm MRE. There were no significant differences in accuracy among the four time-points (T0, 1.41 mm; T1, 1.53 mm; T2, 1.58 mm; T3, 1.47 mm). The pronasale, stomion inferius (Stmi), stomion superius (Stms) showed an increase in MRE (P < 0.01, P < 0.05, and P < 0.05, respectively), whereas the Lower Lip showed a decrease in MRE (P < 0.01). There were no significant differences in errors across time-points for the soft-tissue B point, soft-tissue Pogonion, or soft-tissue Menton between the genioplasty and non-genioplasty groups. Conclusions: The AI algorithm in this study might be an effective tool for STLI in Lat-Cephs at T1, T2, and T3, despite the presence of brackets, S-PS, fixed retainers, genioplasty, and bone remodeling.