Learning to simulate realistic human diffuse reflectance spectra.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Artificial Intelligence

Background:

• Hyperspectral imaging offers clinical potential but lacks efficient methods for relating spectral data to tissue parameters.
• Accurate spectral data is crucial for training and validating artificial intelligence (AI) algorithms in biomedical imaging.
• Current gold-standard Monte Carlo (MC) simulations are computationally too expensive for large-scale use.

Purpose of the Study:

• To develop a scalable and accurate method for generating realistic tissue reflectance spectra.
• To support AI development and validation in biomedical imaging applications.

Main Methods:

• Trained a general-purpose neural surrogate model using over 50 million MC simulations.
• Validated the model against over 5000 in vivo hyperspectral images from open surgery, annotated with 23 tissue classes.
• Assessed clinical potential by evaluating the recovery of organ-specific oxygenation dynamics in a porcine model.

Main Results:

• The surrogate model achieved MC-level accuracy with significantly faster inference (five orders of magnitude).
• Improved spectral recall by 13-48 percentage points over existing models across 140 million human tissue spectra.
• Demonstrated suitability for recovering organ-specific oxygenation dynamics in a controlled experiment.

Conclusions:

• Neural surrogate models can provide MC-level accuracy and in vivo realism at minimal computational cost.
• Enables large-scale, efficient data generation for biomedical optics and robust AI development for clinical applications.