Deep learning-based pigment analysis model trained with optical approach and ground truth assistance

Area of Science:

• Dermatology and biomedical optics
• Medical imaging and artificial intelligence
• Computational biology and bioinformatics

Background:

• Deep learning is widely used for skin pigment analysis (melanin, hemoglobin).
• Traditional regression methods face limitations in input image resolution due to computational constraints.
• Existing optical approaches can compromise performance when addressing resolution limitations.

Purpose of the Study:

• To develop an integrated training method for skin pigment analysis that overcomes image resolution limitations while preserving performance.
• To enhance the accuracy of melanin and hemoglobin quantification in skin images.
• To enable applications such as simulating treatment effects through pigment modification.

Main Methods:

• An integrated training strategy combining an optical approach with ground truth data.
• Decomposition of skin images into melanin, hemoglobin, and shading maps.
• Reconstruction of pigment maps by solving the forward problem using ground truth references.

Main Results:

• High correlation coefficients achieved against the VISIA system: 0.978 for melanin and 0.975 for hemoglobin.
• The model successfully overcomes image resolution limitations without performance compromise.
• Demonstrated capability in producing pigment-modified images for treatment simulation.

Conclusions:

• The proposed integrated training method offers a robust solution for accurate skin pigment analysis.
• This approach enhances the utility of deep learning in dermatology and cosmetic applications.
• The model's ability to simulate treatment effects opens new avenues for personalized skincare and medical interventions.