356
M.F. Wilkins . L. Boddy . G.B.J. Dubelaar
likely category for the data pattern. ANNs have the advantage in comparison to
multivariate statistics of making no prior assumptions as to the nature of the
category data distributions, and once trained, they are fast and efficient in use. In
most of the early studies applying ANNs to AFC data only a few taxonomie
categories were discriminated (e.g. Frankel et al. 1989, 1996; Morris et al. 1992;
Balfoort et al. 1992; Smits et al. 1992; Wilkins et al. 1994, 1996). Scaling up is
not a trivial task, but 36 to 72 phytoplankton species, grown in artificial culture,
have now been discriminated (with 70% overall successful identification) using
multilayer perceptron (MLP) and radial basis function (RBF) ANNs (Boddy et al.
1994, 2000; Wilkins et al. 1999). The latter offers significant advantages in the
ability to reject data patterns not corresponding to any of the classifications known
to the network (Morris and Boddy 1996, Wilkins et al. 1999). While some species
are always identified with high success, others are not, due to overlap of character
distributions. To improve discrimination, additional and/or different
discriminatory characters are required.
Conventional flow cytometers use analog data capture electronics to collect
summary statistics for each pulse, such as pulse width, peak pulse height and
integrated value. However, the pulse shape of the light scatter and fluorescence
signals, acquired as the analysed particle traverses the beam focus, may weIl
contain additional discriminatory information wh ich can be exploited (Godavarti
et al. 1996). CytoBuoy, an autonomous AFC designed for mounting within a buoy
for in situ sampling (Dubelaar et al. 1999; Dubelaar and Gerritzen 2000) uses a
single green (532 nm) laser and retains full digital pulse shape information for
four signals: forward scatter (FSC), side scatter (SSC), orange fluorescence (FLO)
and red fluorescence (FLR), sampled at 4MHz. The raw 8-bit sampie values for
each signal are accumulated in an internal 64kB data buffer before transmission to
the shore by radio link; this limits the number of partieles in any one sampling run
to a few thousand, depending on particle size.
The CytoBuoy is designed to process a relatively wide sampie stream. To
maintain a sufficiently large depth of focus over this stream, the laser focus width
cannot be reduced to less than 5/Lm. The measured pulse shapes are the
convolution of the light intensity profile across the laser focus with the particle
shape (or distribution of optically emitting material) along its longest axis (as
particles flowing through the laser beam are stretched by the fluid acceleration).
The measured pulse shapes of partieles smaller than about 5/Lm are essentially
dominated by the Gaussian shaped light intensity profile, with litde or no
influence from the particle shape itself. The shape features of particles
bigger/longer than a few times the laser focus width (about 20 /Lm and larger) are
weIl expressed in the detector pulses (Fig. 18.1a), whereas for intermediate sized
particles (roughly 5 - 20 /Lm) the shape expression varies from little to reasonable.
In this paper neural net analysis of CytoBuoy pulse shape data is investigated.
In order to apply ANNs to pulse shape analysis, the first step is that of feature
extraction: converting the raw pulse representation (i.e. a variable length sequence
of sampie values) into a more compact representation capable of capturing the
variation between the different categories (i.e. a small number of characteristic
measurements, each quantifying a different aspect of the pulse shape). In addition
it is desirable, if possible, to separate out information on particle size and overall
M.F. Wilkins . L. Boddy . G.B.J. Dubelaar
likely category for the data pattern. ANNs have the advantage in comparison to
multivariate statistics of making no prior assumptions as to the nature of the
category data distributions, and once trained, they are fast and efficient in use. In
most of the early studies applying ANNs to AFC data only a few taxonomie
categories were discriminated (e.g. Frankel et al. 1989, 1996; Morris et al. 1992;
Balfoort et al. 1992; Smits et al. 1992; Wilkins et al. 1994, 1996). Scaling up is
not a trivial task, but 36 to 72 phytoplankton species, grown in artificial culture,
have now been discriminated (with 70% overall successful identification) using
multilayer perceptron (MLP) and radial basis function (RBF) ANNs (Boddy et al.
1994, 2000; Wilkins et al. 1999). The latter offers significant advantages in the
ability to reject data patterns not corresponding to any of the classifications known
to the network (Morris and Boddy 1996, Wilkins et al. 1999). While some species
are always identified with high success, others are not, due to overlap of character
distributions. To improve discrimination, additional and/or different
discriminatory characters are required.
Conventional flow cytometers use analog data capture electronics to collect
summary statistics for each pulse, such as pulse width, peak pulse height and
integrated value. However, the pulse shape of the light scatter and fluorescence
signals, acquired as the analysed particle traverses the beam focus, may weIl
contain additional discriminatory information wh ich can be exploited (Godavarti
et al. 1996). CytoBuoy, an autonomous AFC designed for mounting within a buoy
for in situ sampling (Dubelaar et al. 1999; Dubelaar and Gerritzen 2000) uses a
single green (532 nm) laser and retains full digital pulse shape information for
four signals: forward scatter (FSC), side scatter (SSC), orange fluorescence (FLO)
and red fluorescence (FLR), sampled at 4MHz. The raw 8-bit sampie values for
each signal are accumulated in an internal 64kB data buffer before transmission to
the shore by radio link; this limits the number of partieles in any one sampling run
to a few thousand, depending on particle size.
The CytoBuoy is designed to process a relatively wide sampie stream. To
maintain a sufficiently large depth of focus over this stream, the laser focus width
cannot be reduced to less than 5/Lm. The measured pulse shapes are the
convolution of the light intensity profile across the laser focus with the particle
shape (or distribution of optically emitting material) along its longest axis (as
particles flowing through the laser beam are stretched by the fluid acceleration).
The measured pulse shapes of partieles smaller than about 5/Lm are essentially
dominated by the Gaussian shaped light intensity profile, with litde or no
influence from the particle shape itself. The shape features of particles
bigger/longer than a few times the laser focus width (about 20 /Lm and larger) are
weIl expressed in the detector pulses (Fig. 18.1a), whereas for intermediate sized
particles (roughly 5 - 20 /Lm) the shape expression varies from little to reasonable.
In this paper neural net analysis of CytoBuoy pulse shape data is investigated.
In order to apply ANNs to pulse shape analysis, the first step is that of feature
extraction: converting the raw pulse representation (i.e. a variable length sequence
of sampie values) into a more compact representation capable of capturing the
variation between the different categories (i.e. a small number of characteristic
measurements, each quantifying a different aspect of the pulse shape). In addition
it is desirable, if possible, to separate out information on particle size and overall