A comparative analysis of artificial intelligence generated 3d facial models from 2D photos: A procrustes-based validation of landmark accuracy in Orthodontics.

Objectives: The primary aim of this study was to evaluate the accuracy and precision of a deep learning artificial intelligence (AI) model in reconstructing 3D facial geometry from a single 2D photograph. The study specifically measured the coordinate differences of human-placed facial landmarks on the AI-generated models against those placed on 3D models from a gold-standard scanner (Bellus 3D). Secondary objectives included identifying dimensional variations (X, Y, Z) and systematic biases in the AI reconstructions. Methods: This was a retrospective comparative study using a sample of 47 anonymized patient records from the University of the Pacific database. For each patient, a 3D facial model was generated from a 2D photograph using the 3DDFA-V2 deep learning model. These AI-generated models were compared to reference models captured on the same day using a Bellus 3D Scanner. Two trained observers identified a set of facial landmarks on both the AI-generated and Bellus 3D models. The coordinate data were normalized using Procrustes analysis to compare shape and landmark position independent of scale and orientation. The differences in landmark coordinates between the two model types were then statistically analyzed. Results: The analysis revealed that human landmark identification on the AI-generated 3D models reconstructed from 2D photos showed consistently lower variability compared to the gold-standard scans, suggesting that those models may lead to more repeatable measurements. The greatest overall deviations (Euclidean error) were found in landmarks associated with the lateral and inferior borders of the face, such as Right Gonion (RGo), Trichion (Tr), and Left Gonion (LGo). Quantitatively, the mean Procrustes-normalized error across all landmarks was 0.19 ± 0.08 units (range: 0.08-0.45), corresponding approximately to 0.2 mm of proportional deviation when referenced to the 1:1 scale Bellus3D models. While the AI-generated model demonstrated high fidelity, systematic biases were observed mainly as inferior displacements of key landmarks, including Menton (-0.35 units) and Pogonion (-0.30 units), reflecting a consistent downward shift in AI-based reconstruction. Conclusions: The 3DDFA-V2 deep learning application demonstrated in this study is a promising low-cost tool for generating 3D facial models from 2D photographs with a high degree of precision in landmark placement. While the AI models facilitated more repeatable landmarking, they also exhibited some systematic biases. These findings have clinical implications for the use AI-generated 3D models in diagnosis and treatment planning, highlighting areas where further refinement of the AI algorithm is needed.