Machine Learning Models for Forward and Inverse Problems in Diffuse Reflectance Spectroscopy

This thesis addresses the challenges encountered in diffuse reflectance spectroscopy (DRS) regarding the accuracy and efficiency of existing forward and inverse models. To overcome these challenges, we propose the application of machine learning methods, utilizing Monte Carlo data and experimental data from lab phantoms, to develop improved models that enhance accuracy, versatility, and time efficiency.

Initially, surrogate models are developed using machine learning techniques and compared to Monte Carlo simulations for forward predictions in terms of accuracy and time efficiency. This enables us to evaluate the effectiveness of machine learning approaches in enhancing predictive capabilities.

Furthermore, a transfer learning-based model is developed to calibrate the system and improve accuracy. By leveraging pre-trained models, this approach significantly reduces the calibration effort while maintaining high precision.

In the context of inverse modeling, we modify the traditional Monte Carlo lookup table approach by replacing the forward model with a trained neural network. This neural network-based inverse model is then validated using DRS phantom datasets, confirming its effectiveness in accurately estimating optical parameters.

Finally, our aim is to develop a novel neural network-based inverse model that eliminates the need for generating a lookup table. By directly utilizing the neural network, this model not only enhances accuracy and efficiency but also opens pathways for the development of superior clinical diagnostic tools empowered by machine learning.

The outcomes of this thesis contribute to advancing the field of DRS by overcoming existing limitations in accuracy and efficiency. By harnessing the power of machine learning, our proposed models offer improved predictive capabilities, calibration efficiency, and accurate estimation of optical parameters. These advancements lay the foundation for the development of next-generation clinical diagnostic tools with superior performance and reliability.