Comparison of individualized facial growth prediction models based on the artificial intelligence and partial least squares – with longitudinal growth data from Mathews growth collection

Introduction: To develop facial growth prediction models using artificial intelligence (AI) under various conditions, and to compare the performance of these models with each other as well as with the partial least squares (PLS) growth prediction model. Materials and Methods: Longitudinal lateral cephalograms from 33 subjects in the Mathews growth collection were utilized. The dataset included 1,257 pairs of before and after growth lateral cephalograms. In each image, 46 hard and 32 soft tissue landmarks were manually identified. Growth prediction models were constructed using both a deep learning method based on the TabNet deep neural network and PLS method. The prediction accuracies of the two methods were compared. Results: On average, AI showed 0.61 mm less prediction error than PLS. Among the 77 predicted landmarks, AI was more accurate than PLS in 60 landmarks. When comparing AI models with varying numbers of training epochs, those with higher epochs yielded more accurate predictions. Overall, both PLS and AI methods exhibited greater prediction errors for soft tissue and mandibular landmarks compared to hard tissue and maxillary landmarks. However, the AI method showed a smaller increase in prediction error in areas with greater variability. Conclusions: AI proved to be a valuable growth prediction method, with clinically acceptable prediction errors averaging 1.49 mm for 45 hard tissue landmarks and 1.71 mm for 32 soft tissue landmarks. PLS accurately predicted landmarks with low variability. However, AI generally outperformed PLS, particularly for landmarks in the lower part of the craniofacial structure and soft tissue, where uncertainty is considerable.