Draelos Rachel L, Kesty Chelsea E, Kesty Katarina R
Vismedica AI, LLC, Durham, North Carolina, USA.
St. Petersburg Skin and Laser, St. Petersburg, Florida, USA.
J Cosmet Dermatol. 2025 Apr;24(4):e70050. doi: 10.1111/jocd.70050.
Due to high patient demand, increasing numbers of non-dermatologists are performing skin assessments and carrying out laser interventions in medical spas, leading to inferior outcomes and higher complications. A machine learning tool that automatically analyzes patient skin has the potential to aid non-dermatologists.
To develop a high-performing machine learning model that predicts Fitzpatrick skin type, hyperpigmentation, redness, and wrinkle severity simultaneously.
We developed the SkinAnalysis dataset of 3662 images, labeled by a dermatologist across five skin scales. We trained and evaluated machine learning models across 15 different configurations, including three neural network architectures and two loss functions.
The best-performing model was an EfficientNet-V2M architecture with a custom cross entropy loss. This model's mean test set accuracy across all labels was 85.41 ± 9.86 and its mean test set AUROC was 0.8306 ± 0.09599. An interesting trend emerged in which machine learning model performance was higher at the extremes of the scales, suggesting greater clinical ambiguity in the middle of the scales.
Machine learning models are capable of predicting multiple skin characteristics simultaneously from color photographs of the face. In the future, similar models could assist non-dermatologists in patient skin evaluation to enhance treatment planning.
由于患者需求旺盛，越来越多的非皮肤科医生在医疗美容机构进行皮肤评估和激光治疗，导致效果不佳和并发症增多。一种能够自动分析患者皮肤的机器学习工具可能会帮助非皮肤科医生。
开发一种高性能的机器学习模型，可同时预测菲茨帕特里克皮肤类型、色素沉着、皮肤发红和皱纹严重程度。
我们开发了包含3662张图像的皮肤分析数据集，由皮肤科医生按照五个皮肤等级进行标注。我们在15种不同配置下训练和评估机器学习模型，包括三种神经网络架构和两种损失函数。
表现最佳的模型是采用自定义交叉熵损失的EfficientNet-V2M架构。该模型在所有标签上的平均测试集准确率为85.41±9.86，平均测试集AUROC为0.8306±0.09599。出现了一个有趣的趋势，即机器学习模型在等级的两端表现更好，这表明在等级中间存在更大的临床模糊性。
机器学习模型能够从面部彩色照片中同时预测多种皮肤特征。未来，类似的模型可以帮助非皮肤科医生进行患者皮肤评估，以改进治疗方案。
