Physiological characterization of skin lesion using non-linear random forest regression model.

Area of Science:

• Dermatology
• Medical Imaging
• Machine Learning

Background:

• Current melanoma diagnosis relies on superficial skin assessment, neglecting vital subsurface information.
• Physiological skin features, like melanin concentration, are critical indicators in melanoma development.

Purpose of the Study:

• To propose a non-linear model for extracting physiological skin features, specifically eumelanin and pheomelanin concentrations.
• To enhance melanoma diagnosis by incorporating subsurface information from standard or dermoscopic images.

Main Methods:

• Development of a random forest regression model for non-linear physiological feature extraction.
• Characterization of eumelanin and pheomelanin concentrations from camera and dermoscopic images.
• Validation through phantom studies and separability tests using clinical images, comparing against linear and non-linear models.

Main Results:

• The proposed random forest regression model demonstrated superior performance compared to existing linear and non-linear models.
• The model accurately characterized eumelanin and pheomelanin concentrations in both phantom and clinical experiments.
• Significant improvements were observed in the separability of melanoma cases using the proposed quantitative feature extraction.

Conclusions:

• The developed model offers quantitative characterization of skin features, providing crucial subsurface information.
• This approach enables dermatologists and clinicians to achieve more accurate and improved melanoma diagnoses.
• The findings highlight the potential of advanced machine learning models in revolutionizing dermatological diagnostics.