Facial Recognition Neural Networks Confirm Success of Facial Feminization Surgery

Sir: We read with great interest the article by Chen et al., “Facial Recognition Neural Networks Confirm Success of Facial Feminization Surgery,” and commend the authors for an innovative and clinically relevant application of an increasingly prevalent technology in our society.1 Facial recognition technology is now routinely used in personal smartphones, transportation hubs, department stores, and border crossings.2 In medicine, facial recognition technology algorithms have successfully been shown to detect congenital facial dysmorphisms and to measure increases in emotional expression of happiness following facial reanimation procedures.3–5 Chen et al. have demonstrated another intriguing application of facial recognition technology for surgical outcomes assessment.1 Using four publicly available facial recognition technologies based on deep neural networks, they evaluated the accuracy with which these algorithms classified the perceived sex of 20 male-to-female transgender patients who underwent facial feminization surgery procedures. Preoperatively, frontal photographs of the patients were classified as female only 53 percent of the time. Postoperatively, across all algorithms, the facial recognition technologies classified frontal photographs of transgender female patients as female 98 percent of the time. This study demonstrates the accuracy with which modern facial recognition technology algorithms can classify human sexes across various ages and, by extension, corroborates the aesthetic success of facial feminization surgery procedures. Several interesting questions are raised by this study that deserve further investigation. First, given a relatively small sample size of 20 patients consisting of 12 Caucasians, four African Americans, two Hispanics, one Asian, and one Native American, the authors did not examine differences in classification confidence score between ethnicities. A larger sample would enable more statistically meaningful conclusions. This is an important consideration, as many current facial recognition technologies are inherently biased based on their training data sets, with algorithms commonly performing more accurately when classifying Caucasians compared to other ethnicities.2 Second, current facial recognition technologies have major performance limitations with respect to lighting, pose, and disguise or camouflage, which may include accessories such as glasses, wigs, or jewelry, as well as makeup.2 Wearing makeup can significantly alter an individual’s appearance. In Figure 3, the subject does not appear to have any makeup in the preoperative image, but they are clearly wearing makeup in the postoperative image.1 To critically evaluate the performance of facial recognition technologies on surgical changes in soft tissue and skeletal structure would necessitate that preoperative and postoperative photographs be taken without the camouflaging effects of makeup, which may be a confounding variable that dramatically impacts facial recognition technology performance. Lastly, the authors do not comment on whether any of these patients underwent hormonal therapy and the timing of treatment for those who did. Hormonal therapy can elicit dramatic physiologic effects on soft tissue and bony structure independent of surgery. The use of hormonal therapy without surgical intervention would be worthwhile to investigate with respect to facial recognition technology performance. As suggested by the authors, future studies could also examine the marginal impact of individual facial feminization surgery procedures on facial recognition technology performance to determine which procedures elicit the greatest effects. We commend the authors on a creative, thoughtful, and well-executed study and eagerly anticipate the expanding applications of facial recognition technology in plastic and reconstructive surgery to evaluate and improve patient outcomes. DISCLOSURE The authors have no financial interest to declare in relation to the content of this communication.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.