Advancing Label Free Imaging for Understanding Tissue Structure and Health

Abstract

A wide range of biomedical imaging technologies are available for clinicians and researchers to improve diagnosis and treatment of disease. This work studies label free imaging systems and image processing with four different focuses. Image science is utilized for each focus in order to maximize the amount and quality of information which can be derived from images. Advancing image processing techniques is particularly appealing for biomedical purposes as imaging internal organs can be costly and invasive. As more information can be derived from images fewer images will be needed to draw confident conclusions, which can lower barriers and broaden horizons for research studies and diagnostics.Understanding the potential and limitations of these imaging techniques is key in guiding the course of future innovations. Smartphone cameras are not typically considered to be a biomedical tool, however they have high potential due to their image quality and ubiquity. Smartphone-based spectroscopy in particular has been investigated by a number of groups. Here, it is implemented in a benchtop system which utilizes a fiber optic cable to take diffuse reflectance spectroscopy measurements from a sample. With this system spectral measurements can be taken with ease. Key challenges emerge when processing the data from raw image captures to calibrated spectra. Among them, establishing a pixel to wavelength relationship and determining the best method to work with the data from different color channels. Polarized light imaging (PLI) is a useful technique for measuring structural elements of certain types of tissue, neural tissue is of particular interest as white matter tracts have been found to be intrinsically birefringent. As a result, PLI has been investigated for microstructural neural mapping purposes which are desirable for a number of applications such as surgical guidance and diffusion magnetic resonance imaging (dMRI) validation techniques. To assess the potential of PLI in a backscattering configuration for these purposes a multispectral polarimeter is used to take measurements of ferret brain samples in regions with fiber structures of interest. From here, to further understand the sensitivity and limitations of backscattering polarimetry, tissue phantoms are used and directly compared with the neural tissue. Relaying image information which can be clearly visually understood by a viewer to relevant quantifiable information is a key challenge in computer vision. Image texture is among the properties which can be visually differentiated with ease but a difficult to quantify. One popular method for doing so is the use of Haralick texture feature extraction which relies on a grey level coocurrence matrix (GLCM) to enumerate all the pairs of pixels with specified grey levels. From here the 14 Haralick features are calculated by applying equations to the GLCM to quantity the distribution of grey levels. This method has been applied in a broad range of applications, however there is often not a lot of discussion about the meaning of the features making them somewhat difficult to interpret and contextualize after being calculated. In order to try to expand on this simulated texture images are produced and the texture feature extraction process is carried out in order to create a visual scale for each feature. This is then related to real data in order to assess the limitations of this method. Endoscopy is a major tool clinicians have at their disposal for detecting and monitoring gastrointestinal disease progression. Typically, white light imaging is used, giving doctors a view of the tissue similar to what they would see if they could directly look at it. However, there are many other imaging types which can be added to endoscopes to expand the visualization beyond that and highlight structural and biochemical changes the tissue is undergoing. In order to assess four different imaging modes, fresh healthy, metaplastic and cancerous esophageal tissue samples were excised from people during upper endoscopy procedures at the University of Arizona Cancer Center. Autofluorescence, hyperspectral, optical coherence tomography and polarized light imaging were used to measure the samples. From here the data could be processed by extracting potentially relevant image features and running them through classification and analysis algorithms in order to determine which of these would be best to incorporate into an endoscope independently or in a multimodal configuration in the future.