Instrumentation & Measurement Magazine 25-1 - 64

heterogeneous deposition of the drop resulting in non-uniform
sampling could be detrimental to the reproducibility of
results [1]. In addition, the data collection process is labor-intensive,
and limits the collection of a large Raman data set that
is required for minimizing the standard deviation, maximizing
the SNR, and applying machine learning techniques. To
overcome these limitations, we have developed a flow-cell
based RS technique [13] that provides highly reproducible Raman
spectral measurements of whole blood or blood lysate
and blood components such as plasma and serum. The method
consists of repeatedly interrogating a flowing blood sample
to acquire and then sum-up a large number of subsequent
RS measurements from multiple sample sites to optimize
the SNR. A common method for enhancing the Raman SNR
is surface enhanced Raman spectroscopy (SERS). This uses a
roughened metal surface or metal nanoparticles ~50 nm in size
to achieve at least a million-fold enhancement of the Raman
signal. Although there are some successful demonstrations
of SERS-based blood biopsy, this technique has several challenges
that impact the reliability and reproducibility of data
[14], and hence it is not considered here.
To ensure measurement reproducibility, the measurement
of Raman spectra of the blood sample should be preceded by
a two-step calibration procedure. This involves collecting the
Raman spectra from standard Raman reference materials such
as silicon, polystyrene, and cyclohexane in order to correct for
spectral shifts. A second step involves measurement of the Raman
spectrum of a fluorescence standard (SRM 2241, National
Institute of Standards and Technology, USA) that allows for
the correction of the instrument response in order to account
for lower sensitivity at higher Raman shifts.
Raman Signal Contaminants and Data
Preprocessing Techniques
Due to the nature of the instrumentation and the process of
measurement, many sources of Raman signal contamination
may potentially arise [15], introducing spectral artifacts that
are not related to the pathological state of the blood. While following
strict measurement protocols may be able to control for
some detrimental factors, many are simply too difficult to account
for in the experimental design. Examples of such factors
include temperature, differences in the operator of the RS system,
or even variations in the day of measurement. Instrument
characteristics can also impact spectral signature with variations
associated with different optical components including
the spectrometer, laser, lenses, and interference filters. Preprocessing
techniques aim to eliminate unwanted and irrelevant
variations, allowing for previously obscured features relevant
to the pathological state to emerge. Table 1 lists examples of the
main Raman signal contaminants in the spectra of blood and
examples of preprocessing techniques used for their removal.
The reader is referred to [4] and [15] for a more comprehensive
list of preprocessing techniques. Fig. 2 illustrates the effect of
the application of these techniques on representative Raman
spectra of blood lysate data reported in [13].
Since large fluorescence backgrounds, and thus large photonic
counts, are typically observed in blood, shot noise-rather
than other sources of stochastic noise including dark noise or
CCD read-out noise-is usually the most significant source of
noise in RS [16]. Shot noise is due to the discrete nature of the detected
stream of photons and can be reduced by increasing data
acquisition times to maximize the occupied fraction of the sensor
dynamical range and by averaging over multiple spectral
accumulations. Another widely used tool for noise reduction
after spectral collection is the Savitzky-Golay (S-G) filter [15].
Savitzky-Golay smoothing is governed by two parameters, the
window size and polynomial order [4]. Smoothed points are determined
by iteratively fitting polynomials across the spectrum.
Increased window sizes increase smoothing but can significantly
reduce Raman peak heights. Therefore, these parameters
must be optimized to compromise between smoothing and information
preservation, i.e., peak detectability.
Due to the weak nature of Raman scattering, long exposure
times of at least a few seconds are often required to acquire a
Table 1 - Main signal contaminants in the RS of blood and examples of
preprocessing techniques used to mitigate their impact on the signal
Source of Raman
signal contaminant
Instrumentation
Changes in instrument spectral response with
time, and differences between different systems
Cosmic rays
Measurement
Background fluorescence from instrument
components and interrogated sample
Light scattering artifacts
Type of Raman signal contaminant
Shot noise
Preprocessing technique
Curve smoothing technique improving peak
detectability, e.g., Savitzsky-Golay filtering
Raman shift calibration, instrument response
correction
Cosmic ray removal algorithm
Background subtraction algorithms, e.g.,
sensitive nonlinear iterative peak (SNIP)
clipping, low-order polynomial fitting, rolling
ball technique, etc.
Multiplicative scatter correction or standard
normal variate normalization
Difference in sample thickness and concentration Vector normalization
64
IEEE Instrumentation & Measurement Magazine
February 2022
Instrumentation & Measurement Magazine 25-1

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-1
