The spectral radiance measured by an imaging spectrometer for a material on the earth's surface has significant dependence on environmental factors such as the illumination environment and the atmospheric conditions. This dependence has limited the success of material identification algorithms that rely on hyperspectral image data without associated ground truth information. An important advantage of hyperspectral data is that the sensor spectral dimensionality typically exceeds the dimensionality of the signature variability for any material of interest. We have shown, for example, that the set of observed 0.4 - 2.5 micrometers spectral radiance vectors for a material on the earth's surface lies in a low-dimensional subspace of the hyperspectral measurement space. This analysis has led to robust algorithms for invariant subpixel image analysis that have been applied to a number of remote sensing applications. Similar computational methods can be applied to biomedical images by introducing variability models for the signatures of interest. We present results for material identification in remote sensing images as well as for the quantification of cell population in 3D brain tissue samples.
Raman chemical imaging microscopy has been proven to be a powerful methodology for analyzing a wide range of solid state materials. For biomedical applications, Raman chemical imaging has been shown to be effective in assessing clinical samples including breast tissue lesions and arterial plaques. With Raman chemical imaging systems based on microscopes, materials can be analyzed with molecular specificity, without labor intensive sample preparation or the use of dyes and stains at diffraction limited spatial resolution (< 250 nm). However, microscopes cannot readily be used to perform in vivo measurements. With the recent development of flexible fiberscope technology, Raman chemical imaging can be applied within remote and confined environments and the potential exists for in vivo use. This manuscript provides the first description of novel Raman chemical imaging fiberscope technology, including data analysis strategies for extracting information from Raman chemical imaging data sets.
Advances in molecular probes and fluorescence labeling require sophisticated spectral analysis. For the analysis of spectral data one frequently uses an RGB system (3D color system), dedicated filters for each n basic spectra (nD), or a spectral analyzer (mD, m>n). We present a method for designing spectral filters that optimally resolve all basic spectra as well as all linear combinations of two basic spectra.
A high-performance birefringent imaging optical spectrometer (BIOS) based on liquid crystal elements is designed, built, and characterized. The result is a remarkably compact and simple system for spectral imaging of 2D scenes, with high throughput (85%), no moving parts, and perfect spatial registration between images. Key results include resolution of 4 nm shifts and demonstration of near diffraction-limited image quality. One special benefit is that the interferometer has a setting at which all wavelengths are transmitted without loss; this `white light' setting is of practical benefit in focusing and other sample handling steps. The signal-to-noise of interferometric systems is derived theoretically and compared against filters-based instruments for various source spectra. Based on this analysis and the demonstrated performance of the BIOS system, it appears well-suited to applications such as discriminating between multiple fluorescent probes.
A high-resolution multi-spectral imaging system provides a powerful tool for research and clinical applications where accurate color acquisition, analysis and/or reproduction is of importance. For example, a spectral analysis of fluorescence activated tumor cells is of interest in the evaluation of photodynamic therapy procedures. In this paper we present a system that is based on a CCD sensor with a spatial resolution of 1024 X 1024 pixels and a liquid- crystal tunable filter. The CCD sensor operates at a maximum framerate of 30 frames per second and the filter which is mounted in the optical path of the imaging sensor has an average bandwidth of 15 nm.
We describe the use of non-focal interferometric cameras for reconstruction of the 4D power spectral density of incoherent sources. We develop a 4D version of the generalized van Cittert-Zernike theorem to establish the Fourier transform relationship between the mutual coherence function and the power spectral density. We present experimental demonstrations of 4D imaging using a rotational shearing interferometer. We discuss limitations of interferometric imaging systems and consider how sensor systems might evolve to combine the stability of focal systems with the algorithmic sophistication and multidimensional capacity of interferometry.
When imaging the backscattered light from turbid tissue using a broadband illumination source, the random scattering of photons within the tissue causes wavelength-dependent optical coupling between pixels. That is, a photon may exit the tissue surface an extended distance away from its entry point. The resulting spectral crosstalk in the detected image can be explained by studying the mean photon path lengths through the tissue. Considering complex tissue geometries with features such as cylindrical vessels, these photons not only travel multiple paths due to wavelength- dependent absorption and scattering, but may also travel through multiple chromophores. To study the effects of 3D features in object space on backscattered light into the image plane, we have constructed a Monte Carlo simulation capable of modeling 3D photon propagation for a tissue slab with an embedded cylinder. The results of hemoglobin-bearing vessels as a primary chromophore are investigated. Because of the relationship between mean photon path length and photon exit angle, we have shown that the choice of entrance pupil in the imaging system plays an important role on the detected backscatter for the specific case of embedded cylinders.
Scanning laser microscopy is a widely used technique in ophthalmoscopy for providing high-resolution real time images of the retina. We describe a scanning laser ophthalmoscope that acquires retinal images at four wavelengths for the purpose of measuring the oxygen saturation of blood in retinal arteries and veins. Images at all four wavelengths are obtained across a single video frame using a temporal interlacing technique. An extraction procedure then permits analysis of four monochromatic images. A technique for calculating oxygen saturation from a multi-spectral image set is presented, along with preliminary measurements. The choice of wavelengths dramatically affects the oxygen saturation calculation accuracy and we present an optimized wavelength set and the calculated oxygen saturation results. The potential applications for this technology range from the diagnosis of various ophthalmic diseases to the detection of blood loss in trauma victims.
Spectral Imaging (SIm) has dramatically improved our ability to localize and quantitatively analyze multiple nucleic acid targets such as chromosomes, genes and gene transcripts. Studies on metaphase cells such as `Spectral Karyotyping' are less complicated than interphase cell studies because the objects (chromosomes) are spatially separated and ratio- labeled probes can be used to uniquely stain each chromosome type. Our research, however, targets the extensive cytogenetic and phenotypic analysis of interphase cells. The complex organization of interphase chromatin and co- localization of gene transcripts (RNAs) in nuclear or cytoplasmic domains requires unique fluorochrome-labeling for each nucleic acid target. An increasing number of commercially available dyes for probe labeling and software to deconvolute partially overlapping emission spectra has helped to overcome most of these obstacles. This presentation summarizes our experience in analyzing numerical and structural alterations in various human cell types (leukocytes, amniocytes, blastomeres or solid tissue) as well as our approach to multi-gene expression profiling using SIm. Examples illustrate a wide spectrum of groundbreaking techniques for interphase cell analysis. We demonstrate how ten or more chromosomes can be scored in interphase nuclei or the relative level of expression of different transforming RNAs in tumor cells can be measured by SIm.
Tissue samples from colon biopsy have been analyzed using FT-IR microscopy. For differentiation of cancerous versus normal tissue and for imaging of tissue structure linear and quadratic discriminant analysis was applied. A large number of samples has been used for statistical parametrization. After IR mapping, the samples have been analyzed with common pathological techniques. Comparing the pathological classification with the results of the discriminant analysis the resolving power of the later is estimated. The influence of filtering (using derivative spectra, baseline correction, and normalization) has been investigated. In vivo application of IR-spectroscopy using IR laser diodes or grating spectrometers for fast measurements will lead to a reduction of data concerning spectral range and resolution. The accompanying loss in differentiation of tissue is discussed. Images constructed with the discriminant analysis are compared to corresponding images from visible microscopy.
A combination of FTIR spectroscopy and positron emission tomography (PET) is shown to provide new information on tissue. Here we give a first demonstration on the potential of this combination in discriminating tumor tissue from healthy tissue. Examples are taken of cancer grown in muscle tissue in mice. Immediately before thin sections of the cancer tissue were prepared, a radiotracer was injected in the living mouse. Subsequently a native section was immobilized on a CaF2 window and an autoradiographic image was recorded from that immobilized section. FTIR maps of the thin sections were obtained by using an infrared microscopy equipped with computerized XY stage and MCT detector. Principal component analysis was chosen for chemometric evaluation of the spectra. Evaluated data were reassembled into 2D maps and compared with the corresponding PET image.
Near infrared and visible spectroscopic imaging systems were developed which are able to acquire spectroscopic images of samples at a distance, completely without contact. These imaging systems were used to analyze two color test samples and one 15th Century drawing from the Winnipeg Art Gallery. Spectra extracted from the quantitative test sample showed good linearity with ink levels across the visible wavelengths. By using wavelength images which match the wavelength sensitivities of the human eye, color reconstructed RGB images can be created with good color fidelity. Because of its ability to penetrate through some art pigments and inks, near infrared spectroscopic imaging was used to investigate lead-point underdrawings in ancient drawings in order to understand the artistic process better.
Remote sensing techniques now include the use of hyperspectral infrared imaging sensors covering the mid-and- long wave regions of the spectrum. They have found use in military surveillance applications due to their capability for detection and classification of a large variety of both naturally occurring and man-made substances. The images they produce reveal the spatial distributions of spectral patterns that reflect differences in material temperature, texture, and composition. A program is proposed for demonstrating proof-of-concept in using a portable sensor of this type for crime scene investigations. It is anticipated to be useful in discovering and documenting the affects of trauma and/or naturally occurring illnesses, as well as detecting blood spills, tire patterns, toxic chemicals, skin injection sites, blunt traumas to the body, fluid accumulations, congenital biochemical defects, and a host of other conditions and diseases. This approach can significantly enhance capabilities for determining the circumstances of death. Potential users include law enforcement organizations (police, FBI, CIA), medical examiners, hospitals/emergency rooms, and medical laboratories. Many of the image analysis algorithms already in place for hyperspectral remote sensing and crime scene investigations can be applied to the interpretation of data obtained in this program.
We have evaluated and combined the features of three different methods to develop an algorithm for rapidly processing hyperspectral images. The hyperspectra were initially processed with Principal Component Analysis to find the appropriate number of independent components and abstract spectral representations (loadings). Key Set Factor Analysis and SIMPLISMA (SIMPle-to-use Interactive Self- modeling Mixture Analysis) methods were combined to find `pure' wavelengths for the components from the loadings. These `pure' wavelengths were used to product initial guesses for the relative concentrations of the components, and these concentrations were used to predict the pure component spectra. The spectra were further refined by using the method of Alternating Least Squares. The methodology is demonstrated on infrared spectra of a simple, three- component chemical mixture and on a hyperspectral infrared image of cartilage tissue.
FT-IR microspectrometry, particularly in combination with digital imaging techniques shows great promise for in-vivo and ex-vivo medical diagnosis. The statement is based on the knowledge that this method delivers information of the chemical structure and composition of a sample and the fact that any disease is linked to changes in the molecular and structural composition of cells and tissues. Typically, these changes are highly specific for a given tissue structure and are therefore potentially detectable by FT-IR microspectrometry. In this paper we present several approaches for the representation of mid-infrared microspectroscopic data acquired with high spatial resolution by the use of a MCT focal plane array detector. The applicability of image reassembling methodologies like functional group analysis, image reconstruction based on factor analysis and artificial neural network analysis to the IR data is discussed.
Fluorescence microscopy is rapidly becoming a multi- dimensional technique. Many applications generate similar data analysis problems. Whatever the non-spatial dimension (time, energy), users have to make the choice between local analysis and global analysis. For local analysis, the evolution of pixels (or regions of interest) is modeled as a function of the external parameter. Results are displayed as parametric images. For global analysis, multivariate statistical analysis can be used to extract and interpret the significant information (in the presence of redundancy and noise) in the form of eigenimages and eigenfactors. Automatic classification methods start to play a role for the co-location problem, in which pixels are classified into regions corresponding to positive, null or negative correlation. With two or three images, the scatterplot (an estimation of the joint probability density function), can be built. Interactive and automatic correlation partitioning (ICP, ACP) can then be performed. The method we have developed (Parzen estimate of the probability density function followed by the watersheds mathematical morphology approach) does not make assumptions about the shape of clusters. With more than three images, dimensionality reduction must be applied, for visualization purposes and for simplifying classification. This can be done by linear or non-linear methods such as Multi-Dimensional Scaling, Auto-Associative Neural Networks or Self-Organizing Mapping.
Underlying the contrast in a hyperspectral Raman image are complete Raman spectra at each of tens or hundreds of thousands of pixels. Multivariate statistics allows reduction of these large data sets to manageable numbers of chemically significant descriptors that become the image contrast. In most cases an object can be viewed as containing a small number (usually fewer than ten) chemically discrete components, each with its own vibrational spectrum. Principal component analysis (PCA) and exploratory factor analysis (FA) can be used to generate descriptors from the experimentally observed Raman spectra in image data sets. Additionally, PCA and FA can be viewed as optimized weighted signal averaging techniques. FA contrast is generated from all regions of a spectrum that are attributable to one component. The result is better signal/noise ratio than is obtained using the height or area of a single band as image contrast. We will discuss a variety of preprocessing steps such as removing outliers and selecting spectral subregions for data analysis optimization. We will illustrate these concepts using an image of bone tissue.