KEYWORDS: Artificial neural networks, Data modeling, Skin, RGB color model, Education and training, Hyperspectral imaging, Tissues, Nervous system, Performance modeling, In vivo imaging
SignificanceMachine learning models for the direct extraction of tissue parameters from hyperspectral images have been extensively researched recently, as they represent a faster alternative to the well-known iterative methods such as inverse Monte Carlo and inverse adding-doubling (IAD).AimWe aim to develop a Bayesian neural network model for robust prediction of physiological parameters from hyperspectral images.ApproachWe propose a two-component system for extracting physiological parameters from hyperspectral images. First, our system models the relationship between the measured spectra and the tissue parameters as a distribution rather than a point estimate and is thus able to generate multiple possible solutions. Second, the proposed tissue parameters are then refined using the neural network that approximates the biological tissue model.ResultsThe proposed model was tested on simulated and in vivo data. It outperformed current models with an overall mean absolute error of 0.0141 and can be used as a faster alternative to the IAD algorithm.ConclusionsResults suggest that Bayesian neural networks coupled with the approximation of a biological tissue model can be used to reliably and accurately extract tissue properties from hyperspectral images on the fly.
Hyperspectral imaging (HSI) is a powerful tool for noninvasive assessment of skin properties, as it can capture the spectral signatures of different skin layers and components. However, HSI also requires efficient and accurate methods for estimating skin parameters, such as the thickness, scattering, and absorption coefficients of each skin layer, from the measured spectra. In recent years, much research has been done regarding the use of machine learning (ML) methods for reducing the time and computational cost required for estimating parameters, compared to classical methods, such as the inverse Monte Carlo (IMC) or the inverse adding-doubling (IAD) algorithm. In this study, we investigated the impact of using random Fourier features (RFF) with a simple linear regression model, as well as with an artificial neural network (ANN), to estimate parameter values directly from the spectra. We compared the proposed models with the ANN and a 1D convolutional neural network (CNN), both trained using the raw spectra as input. All models were trained on simulated data and evaluated on both simulated and in vivo measured spectra using mean absolute error (MAE). We found that even simple linear regression with RFFs performs comparably to the neural networks trained on raw spectra while having much lower training and inference time. The best results were attained with the RFF-based ANN, having an overall MAE of 0.0226, which is an improvement compared to the 1D-CNN, having an MAE of 0.0284.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.