Fully automated orthodontic photograph analysis by machine learning

The craniofacial anthropometric ratios are very useful in sciences such as dentistry, maxillofacial surgery, developmental studies and plastic surgery. The manual method of analyzing facial photographs requires a lot of time and precision. The aim of this study was to introduce an application tool that fully automates the analysis of facial photographs and compare it with the manual method.

In this cross-sectional study, the database consisted of 395 profile photographs, 271 frontal photographs in smile and 346 frontal photographs at rest. A two-stage fully convolutional network architecture was used for landmark detection. Two methods of manual and automatic analysis were compared in the measurement of 8 variables, including buccal corridor space, ratio of the height of the middle to the lower third of the face, total facial convexity angle, facial convexity angle, nasofacial angle, mentolabial angle, and nasofrontal angle. The agreement between the two methods was evaluated using the paired T-test and intraclass correlation coefficient (ICC). A value of p<0.05 was considered significant.

 For total facial convexity (P=0.005), nasofacial (P=0.001), and nasolabial (p=0.02) angles, the difference between the two methods was significant. However, no significant difference was found between the two methods for facial convexity, mentolabial, nasofrontal, buccal corridor space, and the ratio of the height of the middle to the lower third of the face no significant difference was observed between the two methods. The ICC for all variables was found to be greater than 0.69 except for the nasolabial angle. For most of the measured variables, the accuracy of the automatic method was similar to that of the manual method.

Machine learning has the potential to be used in clinical soft tissue analysis. It offers the ability to perform reliable and repeatable analyses on large image datasets.