AUTOMATED PEAK IDENTIFICATION IN A TOF-MS SPECTRUM - Quantitative Medical Data Analysis Using Mathematical Tools and Statistical Techniques - page 116

Another important factor is that the full peak identication procedure
should be automated as much as possible for high-throughput data han-
dling. An additional advantage of an automated peak detection algorithm
is that it minimizes human interaction and thus eliminates potential bias
introduced by investigators. It has been reported that in an inter-laboratory
investigation conducted by the NIST, the same samples were analyzed by
MALDI in dierent laboratories. It turned out that when comparing exper-
imental results from dierent laboratories, the reduced data showed more
dierences than the raw data in some cases. The dierences in the reduced
data were traced back to detailed decisions that investigators made when
the data were reduced 12;13 . This result highlights the need for adoption of
common, well tested and well understood, automated methods to avoid ad
hoc methods developed in each research group.
During the past few years, various algorithms have been developed for
peak detection. For example, Bryant et al. nd peaks by cross-correlating
the spectrum with a predened peak lineshape 14 ; Gras et al. nd peaks in
a spectrum by comparing a segment of the spectrum with a template which
describes the peak shape and isotopic pattern 15 .
While the above methods require knowledge about the peak lineshape,
Wallace et al. have developed an algorithm based on iterative segmentation
that makes no assumption about peak shape and does not need to smooth
the data before peak detection 12 . Another peak detection algorithm that
does not depend on the peak shape is due to Jarman et al. 16 In their
approach, a spectrum is viewed as a histogram. In regions where there is
no peak, the spectrum is relatively at and the intensity varies around
a constant. Hence, it can be viewed as a histogram for a noisy uniform
distribution. Deviation from this distribution will be considered evidence
of peak presence.
Eorts have also been made to smooth the spectrum to increase the
signal-to-noise ratio before attempting peak detection. For example, Morris
et al. developed an algorithm to detect peaks based on the mean spectrum
of the spectra from an ensemble of similar samples 17 . By averaging spectra
of similar samples, the noise is reduced. The mean spectrum can be further
smoothed by wavelet denoising.
Though all of the above peak identication procedures proposed rea-
sonable ways for nding peaks in a spectrum, none of them addresses the
condence level in peak position and intensity assignments. Because of the
noise in the spectrum, there are always uncertainties associated with the
estimates made. These uncertainties represent the condence level about