Spectrum-based deep learning framework for dermatological pigment analysis and simulation

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning in dermatology aids automated diagnosis but struggles with ground truth preparation and visual focus.
• Spectrum-based methods offer detailed pigment information but have practical system limitations.

Purpose of the Study:

• To develop a spectrum-based framework for training a deep learning model to generate melanin and hemoglobin distribution maps.
• To overcome limitations of manual ground truth preparation by synthesizing output maps into skin images for regression.

Main Methods:

• Acquired spectral data and created pigment distribution maps using the developed framework.
• Synthesized output maps into skin images for regression analysis, eliminating manual ground truth.
• Simulated pigment variations by adjusting pigment levels and evaluating based on absorption properties, ITA, and pigment indices.

Main Results:

• The model generated accurate reflectance spectra and spectral images reflecting pigment absorption.
• Achieved high correlation coefficients for melanin (0.913) and hemoglobin (0.941) distribution maps compared to the VISIA system.
• Simulated images of pigment variations showed proportional correlation with numerical adjustments to pigment levels.

Conclusions:

• The developed model generates pigment distribution maps comparable to specialized clinical equipment.
• Simulated images with adjusted pigment variations demonstrate the model's utility.
• This spectrum-based deep learning approach shows significant promise for future professional-level dermatological diagnostic tools.