Real-time skin chromophore estimation from hyperspectral images using a neural network

Area of Science:

• Medical imaging
• Biophotonics
• Computational dermatology

Background:

• Hyperspectral imaging offers non-invasive in vivo human skin analysis, revealing invisible parameters like oxygenation and melanin.
• Current methods using optical models and optimization algorithms provide detailed skin maps but are computationally intensive, taking hours for full-face images.
• This slow processing hinders immediate preview and quality control of acquired hyperspectral data.

Purpose of the Study:

• To develop a computationally efficient method for analyzing hyperspectral skin images.
• To accelerate the generation of skin parameter maps for real-time preview and quality assessment.

Main Methods:

• Implemented a neural network trained to emulate a complex optimization-based analysis algorithm.
• Trained the neural network on hyperspectral images from 204 patients and their corresponding, pre-calculated skin parameter maps.

Main Results:

• The neural network generates visually faithful skin parameter maps significantly faster than traditional optimization methods.
• Achieved computation times of 2 seconds for 3-megapixel full-face images and 0.5 seconds for 1-megapixel images.
• Demonstrated promising generalization ability with low average deviation on selected areas, including wide-field images.

Conclusions:

• The developed neural network provides adequate, relatively accurate results for previewing skin parameter maps within seconds.
• This AI-driven approach overcomes the computational limitations of previous methods, enabling faster data assessment.