Multichannel fluorescence imaging is becoming an indispensable tool for post-genomic biological research. Most of the techniques being applied, both in vivo and in vitro, tend to require multiple labeling to visualize different events or to probe various aspects of the same subject. Overlapping emission spectra from multiple fluorescent probes complicates the acquisition and accurate analysis of individual labels and corresponding targets. To address this issue, a computational approach was developed to quantitatively unmix overlapping spectra. We initially constructed models of excitation spectra of individual fluorescent imaging agents as a superposition of multiple Gaussian functions. These models were then used to perform a quantitative unmixing of the combined spectra in milieu using a non-linear least squares optimization technique. We present here a reliable methodology to identify and quantify the individual components from multichannel fluorescent signals. Our results can be easily incorporated into any routine multispectral analysis.
Several visible and near-IR fluorescent nanoparticles have recently been developed and commercialized. These nanoparticles are made of organic “non-toxic” materials and contain multiple fluorochromes that are embedded into the core of the nanoparticle. Two different nanoparticles (X-SIGHT 650: Absorption 650nm; Emission 673nm and X-SIGHT 691: Absorption 691nm; Emission 715nm) were used individually or in combination for the purpose. 0, 2.5, 5.0, 7.5, and 10 pmoles of the nanoparticles were dispensed into five wells of each plate.
Clear/black bottomed 96-well plates were used and the final volume in each of the 55 wells were 0.2 mL. The three plates were imaged using a multispectral imaging system. This multimodal system enables high-sensitivity optical imaging with high-resolution digital X-ray to enable quantification and localization of biomarkers in small animal imaging. It has a 29-excitation filter position wheel ranging from 380-830nm enabling a wide range of fluorescent applications.
For this application, we used an exposure time of one minute with 4x4 binning on the sensor. The f-stop was 2.8 and the camera’s zoom lens was set to image a field of view of 63mm. A 750nm emission filter was used for all image capture. A stack of 15 images was captured at excitation wavelengths from 410nm to 690nm.
The first step in the process of unmixing the overlapping signals from the two probes is to measure the excitation spectrum of each probe individually and then construct a numerical model spectrum using a sum of Gaussian functions. This is accomplished with the use of an interactive software tool and takes less than one minute. The models can be saved in a library and used when analyzing subsequent image captures.
Excitation spectra and Gaussian models
Unmixing is accomplished using the Levenberg-Marquardt method to perform a non-linear, least squares fit of the numerical models of the probes to the measured excitation spectrum at each pixel in the raw data cube. This process must be guided by the automatic selection of initial values for the fitting parameters in order to optimize the results. The fit produces two images, corresponding to the amount of each probe that must be used to produce the best fit to the measured spectrum at every pixel of the input. This process can be extended to any number of probes as long as the problem is mathematically well determined (i.e., there must be at least n+1 independent points in the spectrum in order to unmix n separate probes).
The unmixing results are presented in Figure 2. As a test of the validity of the procedure, we created two additional 96-well plates each with a 5x5 grid with the corresponding dilutions for the individual probes. These grids were imaged under the same conditions as the original 96- well plate where the grid had various combinations of both probes mixed together. The comparison in Figure 2 shows that the unmixed images closely match the images of the two probes taken separately, validating our method. It is important to draw attention to the fact that the left panels of Figure 2 are synthetic; they were created by the numerical analysis and were not captured by a camera. The panels on the right in Figure 2 are images taken from separate 96-well plates that have a 5x5 grid of wells with dilutions. The close resemblance of the images demonstrates the success of the unmixing methodology.
The process of capturing the images and performing the unmixing analysis takes only a few minutes. The tools described here can be used routinely to analyze multiple fluorochrome-based micro or macro biological images.
Note: The results were obtained using the KODAK X-SIGHT Imaging Agents and imaged using the KODAK In-Vivo Multispectral Imaging System FX.
References
Marquardt, D.W., 1963, Journal of the Society for Industrial and Applied Mathematics, v. 11, p. 431-441.
Rao V.L. Papineni, Ph.D. is a biochemist in Research and Development, Carestream Molecular Imaging and can be reached at rao.papineni1@carestreamhealth.com.
Douglas O.S. Wood, Ph.D. is the manager of software development,
Carestream Molecular Imaging and can be reached at douglas.wood@carestreamhealth.com. Carestream Molecular Imaging is a division of Carestream Health, Inc., 4 Science Park, New Haven, CT 06511; www.carestreamhealth.com/go/molecular.
Multichannel fluorescence imaging is becoming an indispensable tool for post-genomic biological research. Most of the techniques being applied, both in vivo and in vitro, tend to require multiple labeling to visualize different events or to probe various aspects of the same subject. Overlapping emission spectra from multiple fluorescent probes complicates the acquisition and accurate analysis of individual labels and corresponding targets. To address this issue, a computational approach was developed to quantitatively unmix overlapping spectra. We initially constructed models of excitation spectra of individual fluorescent imaging agents as a superposition of multiple Gaussian functions. These models were then used to perform a quantitative unmixing of the combined spectra in milieu using a non-linear least squares optimization technique. We present here a reliable methodology to identify and quantify the individual components from multichannel fluorescent signals. Our results can be easily incorporated into any routine multispectral analysis.
To continue reading this article, sign up for FREE to
Membership is FREE and provides you with instant access to eNewsletters, digital publications, article archives, and more.
