使用人工智能面部真实性定位（FAL）模型检测被操纵的在线鼻整形图像

文献信息

中文摘要

背景： 公开可获取的术前术后照片会影响患者期望和对手术成功的认知。在线平台托管着数千张图像，但在数字修饰照片在美容手术中日益常见的时代，这些呈现的真实性仍未得到验证。面部真实性定位（FAL）模型可以检测指示数字篡改的细微像素级不规则性。本研究代表了基于AI的真实性检测器在美容手术媒体中的首次大规模应用。
方法： 对RealSelf.com公共图库（2025年10月）中前600张连续的鼻整形术后照片进行分析，未进行额外筛选或选择。FAL通过Anaconda-Python环境在macOS上本地实现，使用预训练的全局和局部权重。热图输出可视化了可能的数字操纵区域。阳性定义为鼻部区域热图激活。特异性在200张从RAW导出且经过最少处理的假定未编辑临床照片上进行估计；敏感性在50张临床、故意扭曲的照片上进行评估。
结果： RealSelf.com上的操纵流行率为19.5%（117/600；95% CI，16.3-22.7%）。验证假阳性率为1.5%（3/200；95% CI，0.51-4.32%）。检测Facetune生成的几何扭曲的敏感性为100%（50/50；95% CI，93.0-100%）。
结论： 基于AI的几何操纵检测在一个主要平台上识别出大约五分之一公开鼻整形照片中的可疑编辑。验证表明对Facetune生成的编辑具有高特异性和敏感性。将自动真实性检查整合到临床摄影和平台工作流程中可能会提高透明度。

英文摘要

BACKGROUND: Publicly available before-and-after photos influence patient expectations and perceptions of surgical success. Online platforms host thousands of images, yet the authenticity of these representations remain unverified in an age where digitally altered photographs are increasingly common in aesthetic surgery. The Facial Authenticity Localization (FAL) model can detect subtle pixel-level irregularities indicative of digital tampering. This study represents the first large-scale application of an AI-based authenticity detector to aesthetic-surgery media.
METHODS: The first 600 consecutive postoperative rhinoplasty photographs in RealSelf.com's public gallery (October 2025) were analyzed without additional filtering or selection. FAL was implemented locally on macOS via an Anaconda-Python environment with pretrained global and local weights. Heatmap outputs visualized probable regions of digital manipulation. Positives were defined as nasalregion heatmap activations. Specificity was estimated on 200 presumed-unedited clinical photographs exported from RAW with minimal processing; sensitivity was assessed on 50 clinical, intentionally warped photographs.
RESULTS: Manipulation prevalence on RealSelf.com was 19.5% (117/600; 95% CI, 16.3-22.7%). The validation falsepositive rate was 1.5% (3/200; 95% CI, 0.51-4.32%). Sensitivity for detecting Facetune-generated geometric warps was 100% (50/50; 95% CI, 93.0-100%).
CONCLUSIONS: AIbased geometric manipulation detection identifies suspicious edits in roughly one in five public rhinoplasty photos on a major platform. Validation suggests high specificity and sensitivity for Facetune-generated edits. Integration of automated authenticity checks into clinical photography and platform workflows may improve transparency.