Ordinarily, when light scatters from a material, it does so without changing wavelength. Nevertheless, approximately one in a million photons interacts with the material is scatters from, either gaining or losing a little bit of energy; a molecule in the sample undergoes a vibrational transition, either gaining or losing energy, and the photon that scatters either loses or gains the same amount. By analysing the wavelength change using a precision spectrometer, users can determine which molecular bonds are present in the sample, and consequently what the sample is made from. This phenomenon is named the Raman effect, as it was both predicted and discovered by Sir Chandrasekhara Venkata Raman, for which he won the Nobel Prize in Physics in 1930. 

A Raman microscope is one which captures a full Raman spectrum at every pixel in an image. Because of the slow process of taking an image and moving to the next location, a topic of research in the lab is finding methods for speeding up the process of capturing Raman maps of a sample.

Selective sampling

One approach to speeding up the mapping process is to only capture Raman data in pixels that are likely to be of interest to the user; other pixels merely contain similar spectra as their surroundings and their spectra can be estimated from that of their neighbours. Rather than simply raster scanning the sample, the microscope moves to different locations throughout the sample in an attempt to find the pixels that contain interesting features. This requires developing a Raman microscope that can change sampling location quickly and accurately, as well as an algorithm that can quickly determine the next best location to sample from.

Two approaches were taken to finding an algorithm that could rapidly calculate the optimal sampling location; one was based on Kriging, and one was a heuristic based on 2D surface fitting. Both required decomposing the already-captured spectra into a scalar value, which was performed using a simple projection onto a basis vector (normally a suitable principal component). It is then possible to estimate a response surface, or 2D interpolation, of these values over the surface of the sample. The tricky bit is then to decide where to sample next. Fortunately, when one performs a fit using Kriging, one obtains an estimate for the uncertainty at each interpolated point; it is simply a matter of picking the point with the most uncertainty to sample next. Alternatively, a more computationally-efficient approach is to simply fit a linear and a cubic spline to the data, and plot the difference between the two interpolated values. This is a crude estimate for the uncertainty, and can be calculated rapidly. The algorithm simply selects the point with maximum uncertainty to sample next.

Fluorescence pre-screening

An alternative approach is to apply a pre-screening process, in order to group pixels into regions with similar composition. A selection of representative pixels can be selected for each region, and then the region's identity determined without the need to sample every pixel. This can be done in a number of ways, but the one selected for this project was to use tissue autofluorescence. Pixels were grouped based on their texture and morphology, before being sampled. The resulting speed increase was several orders of magnitude over conventional raster scanning.