Virtual Hyperspectral Imaging Toward Data-Driven mHealth
Abstract
Hyperspectral imaging is widely used for obtaining optical information of light absorbers (e.g. biochemical composition) in a variety of specimens or tissues in a labelfree manner. Acquiring and processing spectral data using hyperspectral imaging usually requires advanced instrumentation such as spectrometers, spectrographs or tunable color filters, which are not easily adaptable in developing instrumentation for field-based applications. Also, use of only RGB information from conventional cameras is not sufficient to obtain a reliable correlation with the actual content of the analyte of interest. We propose a new concept of ‘virtual hyperspectral imaging’ to reconstruct the full reflectance spectra from RGB image data. This allows us to use only RGB image data to determine detailed spatial distributions of analytes of interest. More importantly, it simplifies instrumentation without requiring bulky and expensive hardware. Using a datadriven approach, we apply multivariate regression to reconstruct hyperspectral reflectance image data from RGB images obtained using a conventional camera or a smartphone. In developing a reliable reconstruction matrix, it is critical to obtain a training data set of the specimen of study under the same optical geometry since the spectral reflectance and absorbance is sensitive to the detection and illumination parameters. We designed an image-guided hyperspectral system that can acquire both hyperspectral reflectance and RGB data sets under the same imaging configuration to minimize any discrepancies in the hyperspectral reflectance data acquired using different optical sensing geometries. In our technology development, a telecentric lens that is commonly used in machine vision systems but rarely in bioimaging, serves as a key component for reducing unwanted scattering in biological tissue due to its highly anisotropic scattering properties, by acting as a back-directional gating component to suppress diffuse light. We evaluate our spectrometer-less reflectance imaging method using RGB-based hyperspectral reconstruction algorithm for integration into a smartphone application for non-invasive hemoglobin analysis for anemia risk assessment in communities with limited access to central laboratory tests.
Degree
Ph.D.
Advisors
Kim, Purdue University.
Subject Area
Optics|Medical imaging|Health sciences|Information Technology|Medicine
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.