Deep Learning-Based Classification System for Facial Pigmented Lesions to Aid Laser Treatment Decisions

The accurate diagnosis of facial pigmented lesions is essential for selecting appropriate treatment strategies and improving patient outcomes. Deep learning algorithms, particularly convolutional neural networks (CNNs), have shown promise in classifying skin lesions but have been underutilized in differentiating facial pigmented lesions, particularly in the context of laser treatment planning. The aim of this study is to develop and evaluate deep learning models using transfer learning to differentiate among five types of facial pigmented lesions and compare their diagnostic accuracy with that of expert and non-expert dermatologists. A dataset comprising 432 high-resolution clinical images of five facial pigmented lesions, melasma, ephelides, acquired dermal melanocytosis (ADM), solar lentigo, and lentigo maligna/lentigo maligna melanoma (LM/LMM), was collected. Images underwent preprocessing, including white balance correction and region of interest (ROI) extraction. Two CNN architectures, InceptionResNetV2 and DenseNet121, were trained using transfer learning. The models' diagnostic accuracies were compared with those of nine board-certified dermatologists (experts) and 11 noncertified dermatologists (non-experts). The InceptionResNetV2 and DenseNet121 models achieved overall diagnostic accuracies of 87% and 86%, respectively. Both models outperformed expert dermatologists, who had a median diagnostic accuracy of 80%, and non-expert dermatologists, who had a median accuracy of 63%. Notably, both models achieved 100% sensitivity in identifying LM/LMM. The developed deep learning models demonstrated superior diagnostic performance compared to dermatologists in differentiating among facial pigmented lesions. These findings suggest that such models have potential clinical applicability in assisting dermatological diagnosis and guiding appropriate treatment strategies.