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

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate diagnosis of facial pigmented lesions is crucial for effective treatment and patient outcomes.
• Deep learning, specifically Convolutional Neural Networks (CNNs), shows potential for skin lesion classification but is underutilized for facial pigmented lesions, especially for laser treatment planning.

Purpose of the Study:

• To develop and evaluate deep learning models using transfer learning for differentiating five types of facial pigmented lesions.
• To compare the diagnostic accuracy of these models against expert and non-expert dermatologists.

Main Methods:

• A dataset of 432 high-resolution images of melasma, ephelides, acquired dermal melanocytosis (ADM), solar lentigo, and lentigo maligna/lentigo maligna melanoma (LM/LMM) was utilized.
• Two CNN architectures, InceptionResNetV2 and DenseNet121, were trained using transfer learning after image preprocessing.
• Model performance was compared to the diagnostic accuracy of 9 board-certified dermatologists and 11 non-certified dermatologists.

Main Results:

• The InceptionResNetV2 and DenseNet121 models achieved diagnostic accuracies of 87% and 86%, respectively.
• Both models surpassed expert dermatologists (median 80% accuracy) and non-expert dermatologists (median 63% accuracy).
• Both models demonstrated 100% sensitivity in identifying LM/LMM, a critical finding for early cancer detection.

Conclusions:

• Deep learning models exhibit superior diagnostic performance compared to dermatologists in classifying facial pigmented lesions.
• These AI models hold significant potential for clinical application in assisting dermatological diagnosis.
• The findings suggest these models can aid in guiding appropriate treatment strategies for various facial pigmented lesions.