Automated facial landmark measurement using machine learning: A feasibility study

Statement of problem: Information regarding facial landmark measurement using machine learning (ML) techniques in prosthodontics is lacking. Purpose: The objective of this study was to evaluate and compare the reliability, validity, and accuracy of facial anthropological measurements using both manual and ML landmark detection techniques. Material and methods: Two-dimensional (2D) frontal full-face photographs of 50 men and 50 women were made. The interpupillary width (IPW), interlateral canthus width (LCW), intermedial canthus width (MCW), interalar width (IAW), and intercommissural width (ICW) were measured on 2D digital images using manual and ML methods. The automated measurements were recorded using a programming language (Python), and a convolutional neural network (CNN) model was trained to detect human facial landmarks. The obtained data from the manual and ML methods were analyzed using intraclass correlation coefficients (ICCs), the paired sample t test, Bland-Altman plots, and the Pearson correlation analysis (α=.05). Results: Intrarater and interrater reliability values were greater than 0.90, indicating excellent reliability. The mean difference between the manual and ML measurements of IPW, MCW, IAW, and ICW was 0.02 mm, while it was 0.01 mm for LCW. No statistically significant differences were found between the measurements obtained by the manual and ML methods (P>.05). Highly significant positive correlations (P<.001) were obtained between the results of the manual and ML methods: (r=0.996[IPW], r=0.977[LCW], r=0.944[MCW], r=0.965[IAW], and r=0.997[ICW]). Conclusions: In the field of prosthodontics, the use of ML methods provides a reliable alternative to manual digital techniques for carrying out facial anthropometric measurements.