Face comparison analysis of patients with orthognathic surgery treatment using cloud computing-based face recognition application programming interfaces

Abstract

Introduction: This study aimed to investigate whether the postoperative change in patients after orthognathic surgery, whose facial aesthetics was affected, led to detectable differences using Microsoft Azure, Amazon Web Services Rekognition, and Face⁺⁺, which were commercially available face recognition systems.

Methods: Photographs of 35 patients after orthognathic surgery were analyzed using 3 well-known cloud computing-based facial recognition application programming interfaces to compute similarity scores between preoperative and postoperative photographs. The preoperative, relaxed, smiling, profile, and semiprofile photographs of the patients were compared separately to validate the relevant application programming interfaces. Patient characteristics and type of surgery were recorded for statistical analysis. Kruskal-Wallis rank sum tests were performed to analyze the relationship between patient characteristics and similarity scores. Multiple-comparison Wilcoxon rank sum tests were performed on the statistically significant characteristics.

Results: The similarity scores in the Face⁺⁺ program were lower than those in the Microsoft Azure and Amazon Web Services Rekognition. In addition, the similarity scores were higher in smiling photographs. A statistically significant difference was found in similarity scores between relaxed and smiling photographs according to different programs (P <0.05). For all 3 facial recognition programs, comparable similarity scores were found in all photographs taken before and after surgery across sex, type of surgery, and type of surgical approach. The type of surgery and surgical approach, sex, and amount of surgical movement did not significantly affect similarity scores in any facial recognition programs (P >0.05).

Conclusions: The similarity scores between the photographs before and after orthognathic surgery were high, suggesting that the software algorithms might value measurements on the basis of upper-face landmarks more than lower-face measurements.