Facial Recognition Neural Networks Confirm Success of Facial Feminization Surgery

Abstract

BACKGROUND: Male-to-female transgender patients desire to be identified, and treated, as female, in public and social settings. Facial feminization surgery entails a combination of highly visible changes in facial features. To study the effectiveness of facial feminization surgery, we investigated preoperative/postoperative gender-typing using facial recognition neural networks. METHODS: In this study, standardized frontal and lateral view preoperative and postoperative images of 20 male-to-female patients who completed hard- and soft-tissue facial feminization surgery procedures were used, along with control images of unoperated cisgender men and women (n = 120 images). Four public neural networks trained to identify gender based on facial features analyzed the images. Correct gender-typing, improvement in gender-typing (preoperatively to postoperatively), and confidence in femininity were analyzed. RESULTS: Cisgender male and female control frontal images were correctly identified 100 percent and 98 percent of the time, respectively. Preoperative facial feminization surgery images were misgendered 47 percent of the time (recognized as male) and only correctly identified as female 53 percent of the time. Postoperative facial feminization surgery images were gendered correctly 98 percent of the time; this was an improvement of 45 percent. Confidence in femininity also improved from a mean score of 0.27 before facial feminization surgery to 0.87 after facial feminization surgery. CONCLUSIONS: In the first study of its kind, facial recognition neural networks showed improved gender-typing of transgender women from preoperative facial feminization surgery to postoperative facial feminization surgery. This demonstrated the effectiveness of facial feminization surgery by artificial intelligence methods. CLINICAL QUESTION/LEVEL OF EVIDENCE: Therapeutic, IV.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 69 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.