文献速递口腔正畸

不同深度学习模型在三维照片面部吸引力评估中的比较

Comparison of deep learning models for facial attractiveness assessment on 3D photos.
PubMed 文献 · J Dent · 2025-04-06

本研究收集1272张中国正畸患者治疗前的三维照片,由7名正畸医生用视觉模拟量表评估面部吸引力,比较ResNet18、ResNet50、ResNet101、VGG-16、VGG-19、Inception-v3、MobileNet-v2和DenseNet121等卷积神经网络模型的预测表现。结果显示DenseNet121在预测准确性和排序相关性上表现最佳,MobileNet-v2预测准确性最差;ResNet18训练时间最短但仍能准确预测,VGG-19模型最大且训练耗时最多。

要点

  • DenseNet121在预测准确性和排序相关性上表现最佳。
  • MobileNet-v2预测准确性最差;ResNet18训练时间最短但仍能准确预测。
  • VGG-19模型最大且训练耗时最多;ResNet101随训练轮数增加未实现性能提升。

核心结论

  • DenseNet121在预测准确性和排序相关性方面表现最佳。
  • 研究分析了模型准确性与训练效率之间的权衡,为不同计算条件下的模型选择提供参考。

研究设计与样本

  • 研究收集1272张中国6-18岁正畸患者治疗前的三维照片。
  • 由7名中国正畸医生组成的小组使用视觉模拟量表评估面部吸引力。
  • 三维照片转换为二维RGB图像后输入ResNet18、ResNet50、ResNet101、VGG-16、VGG-19、Inception-v3、MobileNet-v2和DenseNet121等卷积神经网络模型。
  • 使用均方根误差、平均绝对误差、Kendall秩相关系数τ和Spearman秩相关系数ρ评估模型性能。

主要发现

  • DenseNet121预测准确性最佳,MobileNet-v2预测准确性最差。
  • ResNet18训练时间最短,但仍能准确预测面部吸引力。
  • VGG-19模型最大且训练耗时最多。
  • DenseNet121在模型性能与训练成本之间取得平衡,在排序相关性、验证损失和验证平均绝对误差方面优于其他卷积神经网络。
  • ResNet101随训练轮数增加,在初始阶段后未能实现性能提升。

关键数值与推荐

  • 评估指标包括均方根误差、平均绝对误差、Kendall秩相关系数τ和Spearman秩相关系数ρ。
  • DenseNet121在排序相关性、验证损失和验证平均绝对误差方面优于其他卷积神经网络。

临床怎么用

  • 作者认为深度学习模型的成功应用有助于准确、精确地评估面部吸引力。

局限性

  • 原文未报告局限性

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

OBJECTIVES: Convolutional neural networks (CNNs) have demonstrated remarkable success in orthodontics. This study aimed to evaluate the accuracy and precision of several prominent CNN models for evaluating the facial attractiveness in Chinese orthodontic patients aged 6-18 years. MATERIALS AND METHODS: A total of 1272 three-dimensional (3D) pretreatment photographs of patients were gathered. A panel of seven Chinese orthodontists assessed facial attractiveness using a visual analog scale. After conversion to two-dimensional RGB images, the data were fed into CNN models, including ResNet18, ResNet50, ResNet101, VGG-16, VGG-19, Inception-v3, MobileNet-v2, and DenseNet121. The performance of these CNNs was evaluated using root mean square error (RMSE), mean absolute error (MAE), Kendall's tau (τ), and Spearman's rank correlation coefficient (ρ). RESULTS: DenseNet121 demonstrated the best, and MobileNet-v2 performed the worst in terms of prediction accuracy. ResNet18 required the shortest training time yet was still able to accurately predict facial attractiveness. In addition, VGG-19 was the largest and consumed the most time during training. DenseNet121 successfully achieved a balance between the model performance and the training cost and outperformed the other CNNs in terms of ranking correlation, validation loss, and validation MAE. Meanwhile, ResNet101 failed to achieve a performance improvement after an initial stage as the number of epochs increased. CONCLUSIONS: DenseNet121 exhibited the best performance in terms of prediction accuracy and ranking correlation. Furthermore, the trade-offs was also analyzed between model accuracy and training efficiency, offering insights into model selection under different computational constraints. CLINICAL SIGNIFICANCE: The successful application of deep learning models helps evaluate facial attractiveness accurately and precisely.

来源:J Dent,查看原文。版权归原作者所有。

引用

  1. Hu J, Wang C, Qu Q, et al. Comparison of deep learning models for facial attractiveness assessment on 3D photos. J Dent. 2025. PMID: 40199417. doi:10.1016/j.jdent.2025.105735

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。