Spectral "Unmixing" with
Hyperspectral Imaging

"Ultimately the ability to "unmix" multiple fluorophores depends on the number of data points, spectral resolution, and wavelength range!"

Figure 1. Spectral emission profiles of Alexa 555, 569, and 594. Emission spectra of each fluorophore strongly overlaps the others.
Figure 2. Tissue section stained with Alexa 555, 569 and 594. The distribution of each fluorophore is clearly segmented**.
  • Linear/nonlinear unmixing: PARISS uses Spectral waveform cross correlation analysis (SWCC) that is not predicated on either linear or non linear "mixing". See "Use and Abuse" section below. This algorithm can only be used with an adequate number of wavelength data points to ensure non-aliased spectral characterization.


  • Tolerance to low S/N: SWCCA provides greater tolerance to weak signals and poor S/N ratio. It may be possible to generate an accurate spectral topographical map (spectral image) where a conventional digital image of the FOV can be low quality (noisy with low contrast).

  • Multiple spectral segmentation: Accommodates up to 15 fluorophore or chromophore profiles simultaneously


  • Libraries: PARISS classifies spectra either automatically or manually and inserts selected spectra into libraries

  • Pseudo-color coding: All classified spectra will be pseudo-color coded by the operator or automatically by the program. The presence of a color code painted on a grayscale spectral image indicates the presence of a correlated class of library spectrum.

  • Thresholding: All spectra presented by the FOV for classification must meet a user determined threshold before it will be painted to a gray scale image

  • Search for abnormalities: Algorithm uses the logical operator "not" so that any area of a FOV that fails to present spectra that correlate to expectations will be specifically called out. Clicking on any such area will return the actual spectrum. The operator can then decide whether or not to add this spectrum to a library.

Use and abuse of the term "unmixing"
 Warning here comes a rant!
In the physical sciences and remote Earth sensing community the term spectral "unmixing" assumes that there are two or more spectral signatures commingled within a single pixel in the field of view (FOV). This does not necessarily mean that objects with independent spectra are intimately commingled and collocated.

For example, the a leaf on a tree may present areas which are normal, diseased, and attacked by insects all in the same pixel in the FOV. Assuming that there is no mist of fog in the atmosphere to create light scattering, then it is safe to say that the combined spectral signature will be a linear mixture of each spectral component.

If these conditions are the same in the biological samples you study then linear unmixing algorithms can work; but for most of us these conditions cannot be assumed. The level of spectral complexity grows exponentially when emitters are truly collocated, and rigorously commingled. In biological samples, spectral signatures can and do change as a function of localized differences in pH, bonding, chemical interactions, ion changes, binding, sterioisomerism, concentration, and physiology.

Some vendors to the life science community began applying the term "unmixing" to the ability to differentiate between overlapping spectra profiles. The plots shown in Figure 1 are perfect examples. So what does it matter? Well, it has a lot to do with your tolerance for "hype".

For example, when multiple fluorophores are spatially separated it is really not that difficult to differentiate between a very large number of seriously spectrally overlapping signatures. PARISS can generate spectral topographical maps of up to 15 fluorophores (actually more; but it is not easy for the human eye to differentiate between more than 15 pseudo-colors).

Insofar as background can be heterogeneous, and typically always present, SWCCA is far better able to accommodate local spectral variations than algorithms that "take a background sample" and assume that this will work over an entire FOV.

The problem with unmixing more than two truly commingled spectral signatures is a significant challenge. In a low signal to noise environment problems escalate even further, especially when there are considerable difference in signal intensity between spectral components.

Most scientists agree that we live in a non-linear universe in which linearity is a special case. It is puzzling therefore, that there are those who assume that linear mixing can occur in a reactive "biological soup". The reality is that all cell components can and will "react" under the influence of its environment; and in the process change its spectral profile.

Yet we hear companies touting their "linear unmixing" apparently suggesting that one algorithm fits all! SWCCA has its pros and cons; but at least it is not based on false assumptions.

In all probability. the real reason behind the use of linear unmixing is because most spectral imaging systems do not generate enough wavelength data points to permit non-linear unmixing algorithms. In other words they do the best they can with what the system is capable of delivering.

If all commercial spectral imaging systems generated enough data it is a fair assumption that most if not all would provide more advanced mathematical algorithms.

Final message - be suspicious when you hear the term "unmixing" and note the context in which it is being used.

_____________________________________________

Go to top

 
Address
  
© 1990-2010 LightForm, Inc all rights reserved
Privacy Policy ............. Disclaimers

* "Alexa Fluor(R)" and "Molecular Probes(TM)" are trademarks of Invitrogen/Molecular Probes
** Taken from "Abstract_FRET_FLIM_Spectral Imaging2" Click here for the pdf