Image-Based Monitoring System for Green Algal
Haematococcus pluvialis (Chlorophyceae) Cells during Culture
Shinsuke Ohnuki1,4, Satoru Nogami1,2,4, Shuhei Ota2,3, Koichi Watanabe2,3, Shigeyuki Kawano2,3 and
Yoshikazu Ohya1,2,*
1
The green microalga Haematococcus pluvialis accumulates
the red pigment astaxanthin accompanied by morphological
changes under stress conditions, including nutrient depletion, continuous light and high temperature. To investigate
the physiological state of the algal cells, we developed the
digital image-processing software called HaematoCalMorph.
The software automatically outputs 25 single-cell measurements of cell morphology and pigments based on color,
bright-field microscopic images. Compared with manual
inspection, the output values of cell shape were reliable
and reproducible. The estimated pigment content fits the
values calculated by conventional methods. Using a random
forests classifier, we were able to distinguish flagellated cells
from immotile cells and detect their transient appearance in
culture. By performing principal components analysis, we
also successfully monitored time-dependent morphological
and colorimetric changes in culture. Thus, combined
with multivariate statistical techniques, the software
proves useful for studying cellular responses to various conditions as well as for monitoring population dynamics in
culture.
Keywords: Astaxanthin Digital image
Haematococcus pluvialis Morphology.
processing
Abbreviations: astaxanthin, 3,30 -dihydroxy-b,b0 -carotene-4,40 dione; CCD, charge-coupled device; CPU, central processing
unit; DMSO, dimethylsulfoxide; FDR, false discovery rate; GUI,
graphical user interface; JRE, Java runtime environment; JPEG,
Joint Photographic Experts Group; LD, periodical light; LL,
continuous light; PC, principal component; PCA, principal
components analysis; PNG, Portable Network Graphics;
RGB, red, green and blue; SCCAP, Scandinavian Culture
Collection of Algae and Protozoa; SVM, support vector machine; TIFF, Tagged Image File Format.
Introduction
Techniques
Laboratory of Signal Transduction, Department of Integrated Biosciences, Graduate School of Frontier Sciences, University of Tokyo,
Kashiwa, 277-8562 Japan
2
JST-CREST, Kashiwa, 277-8562 Japan
3
Laboratory of Plant Life System, Department of Integrated Biosciences, Graduate School of Frontier Sciences, University of Tokyo,
Kashiwa, 277-8562 Japan
4
These authors contributed equally to this work.
*Corresponding author: E-mail, [email protected]; Fax: +81-4-7136-3651.
(Received July 3, 2013; Accepted September 3, 2013)
Haematococcus pluvialis Flotow [a synonym of H. lacustris
(Girod-Chantrans) Rostafinski] is a unicellular green alga
(Chlorophyceae) that has several unique characters. This alga
forms flagellated green motile cells under favorable growth
conditions. When the conditions become unfavorable, for example under nutrient depletion, continuous light or high temperature, the cells become spherical with no flagella (reviewed
by Boussiba 2000). Simultaneously, the alga turns from green to
red to form cysts (resting cells) that have a thick and heavy
resistant cell wall. This color change is due to continuous
resorption of the green pigment Chl and gradual accumulation
of the red pigment astaxanthin (3,30 -dihydroxy-b,b0 -carotene4,40 -dione) (Kobayashi et al. 1997b, Boussiba et al. 1999,
Wayama et al. 2013). Because astaxanthin possesses high antioxidant activity (Kurashige et al. 1990, Palozza and Krinsky
1992, Naguib 2000), its accumulation is considered a survival
strategy against stressful conditions, under which reactive
oxygen species are produced (Boussiba 2000). Accordingly,
astaxanthin-rich cysts are more tolerant to excessive reactive
oxygen species than vegetative cells (Kobayashi et al. 1997a,
Li et al. 2010).
Because cell morphology and pigment content of the algal
cells change during culture (Kobayashi et al., 1991, 2001), it is
important to measure these cellular features in a physiological
study. To monitor the amount of pigments, quantification
with HPLC and spectroscopy (Yuan et al. 1996, Lababpour
and Lee 2006, Ranga et al. 2009) and neural network-based
estimation from cell images (Kamath et al. 2005) have been
used. However, because the values obtained using these methods are volumetric concentrations or averaged cellular contents, the contents in an individual cell cannot be estimated.
Recently, Raman resonance microscopic technology has been
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126, available online at www.pcp.oxfordjournals.org
! The Author 2013. Published by Oxford University Press on behalf of Japanese Society of Plant Physiologists.
All rights reserved. For permissions, please email: [email protected]
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
1917
S. Ohnuki et al.
used to monitor the pigment content in a single cell, but this
requires expensive instruments and long measuring times
(Collins et al. 2011, Kaczor and Baranska 2011, Kaczor et al.
2011). No techniques for quantifying cell morphology in H.
pluvialis exist.
Digital image analysis has been used to characterize cell
morphology with a large set of quantitative features for other
algal cells such as in taxonomic microalgae recognition in samples from an aquatic ecosystem (Walker et al. 2002, Rodenacker
et al. 2006, Mosleh et al. 2012). However, morphological
changes of a particular algal species have not been studied
quantitatively and statistically. In the research area of fermentation engineering, we applied multivariate statistical techniques for analysis of time-dependent changes of brewing
yeasts during fermentation processes (Ohnuki et al. 2013)
after quantifying a large number of morphological features
(Ohya et al. 2005). Particular effort was devoted to the development of robust statistical methods for providing insight into
the dynamic changes in the yeast population (Ohnuki et al.
2012, Ohnuki et al. 2013). Similar statistical approaches were
potentially applicable to monitor the physiological changes of
algal cells.
Here, we present a novel system for monitoring H. pluvialis
cells during culture. To quantify cell morphology and pigments
in H. pluvialis, we developed image-processing software called
HaematoCalMorph. In addition to the algal cells, the software
recognized dead colorless cells due to infection by chytrid
fungus-like organisms (Hoffman et al. 2008) or other causes.
With a combination of multivariate statistical techniques, it
was possible to classify the cells into specific groups reflecting
the life cycle and to monitor dynamic changes in culture.
Results
Development of image-processing software
To quantify several features of the algal cells, we developed
software named HaematoCalMorph. Fig. 1 summarizes the
flow of the analysis composed of five processing steps (for a
detailed algorithm, see the Materials and Methods). As the first
step, color information was pre-processed from bright-field
microscopic images (Step 1). Grayscale images of red, green
and blue channels were decomposed from a color image, and
the background intensities of images from the three channels
were equalized. Then, we detected edges to recognize the
boundary of segments (Step 2). ‘Local edge’ was first detected
by applying the Canny algorithm (Canny 1986) to the red channel images. ‘Global edge’ was detected by tracing bright regions
of the binarized image of high-chroma regions. We then combined these edges to segment cellular regions. The next important and unique step was to distinguish the foreground from the
background (Step 3). To judge whether the segmented regions
were foreground or background, we focused on the average and
variance of intensities. Because background regions are uniformly transparent, the intensity in the region will have high
1918
average and low variance values. However, since foreground
regions are opaque and inhomogeneous, the intensity will
have low average and high variance values. Foreground regions
were further classified into algal cell, colorless cell and non-cell
regions (Step 4). To distinguish non-cell regions, we used chordiogram distance (Toshev et al. 2010) whereby a circular object
was used as a reference. ‘Cell regions’ were then classified as
algal cells or colorless cells using the color information. Finally,
we calculated the morphological parameters (Step 5). For each
categorized region, quantitative values regarding cell morphology and pigments were calculated. The software also outputs
averages of these values in an image or a set of images obtained
from one experimental trial.
The software was written in the Java language and was able
to run on Java SE 6 on Mac OS X 10.8 and Windows 7. RGB (red,
green, blue) color microscopic images 2,0401,536 pixels in size
of H. pluvialis cells were used as test images.
Capability and specification of HaematoCalMorph
To analyze H. pluvialis cells, bright-field microscopic images
were obtained with a color charge-coupled device (CCD)
camera. Fig. 2 shows a flowchart of the image analysis with
HaematoCalMorph. The input folder is composed of subfolders
containing many color images of the cells. The software first
processes the color images one by one and outputs analyzed
images and a text file of single-cell data in the same subfolder in
the output folder. The single-cell data consist of measurements
that reflect cellular morphology and pixel intensity. Then the
software outputs two other text files, one containing data summarized by image and one containing data summarized by the
subfolder in the output folder. A description of the measurements is given in Table 1. The software requires a Java runtime
environment (JRE) and images in the joint photographic experts group (JPEG), tagged image file format (TIFF) or portable
network graphics (PNG) format without size limitation.
We measured the calculation time by analyzing 320 JPEG
images (2,0401,536 pixels). When a computer running JRE
version SE6 on Mac OS X 10.7 with a four-core (eight threads)
Core i7 central processing unit (CPU) (2.8 GHz) and 20 Gbyte
memory was used, it took an average of 105 s to analyze one
JPEG image. It was equivalent to 30 s for images with four cells
and 259 s for images with 36 cells, indicating that the time
varied depending on the number of cells in the image. When
the computer had multithread CPUs, the software analyzed
images in different subfolders in the input folder in parallel,
reducing the total calculation time. Using the same computer,
it took 9.3 h to analyze the 320 JPEG images in a single folder,
whereas it took only 1.5 h when the images were in 16 subfolders each containing 20 images.
Reliability and reproducibility of the data
calculated by HaematoCalMorph
To determine the reliability of the software, we compared its
output with the results of the manual analyses from different
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
Image-processing software for microalgae
Step 1
Preprocessing
Blue Channel
B
Input Image
A
Step 2
Segmentation
Step 3
Foreground Detection
Global Edge
F
Dark Regions
I
Step 4
Classification
of Cell Region
Green Channel
C
Red Channel
D
Local Edge
E
Boundary
of Segments
G
Step 5 Calculation of parameters
Mean &
Variance of
Intensity
H
Foreground
Regions
J
Circularity of
Regions
K
Cell Regions
L
Fig. 1 Schematic illustration of image-processing procedures used by HaematoCalMorph. The image processing consisted of the five indicated
steps as described in the Results and the Materials and Methods. Images A–L represent: (A) input; (B) the blue channel; (C) the green channel;
(D) the red channel; (E) local edges detected using the Canny method; (F) high-chroma regions for detection of global edges; (G) boundary of the
segments derived from E and F; (H) regions with high local intensity variances and/or low mean intensities; (I) regions with low intensity; (J)
foreground regions derived from H and I; (K) circular regions based on chordiogram distances; (L) region-type classification and final output. In
image H, the red and green intensities of the segmented regions indicate the mean and the extent of variance of intensity, respectively. In image I,
white and black regions indicate dark and bright regions, respectively, in the combined grayscale image of each of the color channels (B–D).
In image K, black, red and green regions indicate the background regions, non-round regions (chordiogram distance >0.5) and round regions
(chordiogram distance <0.5), respectively. In image L, light blue, red, pink and green lines indicate algae, colorless, touch and other classification
types, respectively.
methods. Of 428 cell areas identified by HaematoCalMorph,
412 areas were coincident with the cell areas inspected by
eye. In addition, the remaining 16 areas were not classified as
algae, implying that the recognition was accurate. Then we
examined whether the software classified cell type properly.
The software classified cell areas into four types, namely
algae, colorless cells, ‘touch’ and other. The ‘touch’ category
included cell areas that contacted the external margin of the
image. We found that the system had excellent positive predictive value (99%) for identifying algae (242 of 245 cells). It
judged 23 of 31 cells correctly as colorless dead cells. No cells
were misidentified between algae and colorless cells. These
results suggest that the classification ability of our system is
highly reliable.
Then we tested the accuracy and reproducibility of two
morphological measurements: the long axes and short axes of
the cells. Three different people manually measured these features three times using 10 algal cells in a photographed image
(Fig. 3, colored crosses). As expected, the results differed individually, with failure to obtain reproducible results. In contrast,
we obtained reproducible data with the software (Fig. 3, analyzed image and diamonds in plots); the outputs were almost
the average of the manual counts or close to them. In some
cases (Cell IDs #6 and #10), manually obtained measurements
were under-represented. In addition, it took 17 min to measure these by hand but only 1.3 min with the software.
Estimation of cellular pigment contents
We examined whether we could obtain accurate information
on pigment content. We applied multiple linear regression analyses to predict cellular Chl and carotenoid content. We cultured cells under continuous light (LL) conditions and obtained
time-course samples. Volumetric Chl and carotenoid concentrations were determined with conventional spectrometric
methods, and cellular pigment contents were calculated by
dividing by the cell density. Using the same samples, microscopic cell images were acquired and analyzed with
HaematoCalMorph to obtain quantitative color data. Since
the cellular pigment contents varied at different time points,
we used these data to generate regression equations to explain
cellular pigment contents with three HaematoCalMorph
parameters, AlgaeInnerMeanRedIntensity, AlgaeInnerMeanGreenIntensity and AlgaeInnerMeanBlueIntensity (Supplementary Table S1). When the regression equation was
applied to the same data obtained under LL conditions, the
predicted values almost completely matched the values determined by the conventional spectrometric method (Fig. 4, left
panels; R2 = 0.999 for Chl; R2 = 0.992 for carotenoids).
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
1919
S. Ohnuki et al.
Fig. 2 Schematic drawing of the analysis flow. Rounded squares indicate folders stored in a computer. A folder containing color images of algal
cells was specified as the input folder. The output consisted of analyzed images and text files. Subfolders in the output folder contained analyzed
images and a text file with single-cell values. Text files outside the subfolders in the output folder contained summarized data organized by image
and by subfolder. For more information and quantitative values in the output text files, see the Materials and Methods and Table 1.
To determine the generality of the method, we applied the
same equation to the different images of cells cultured under
periodic light (LD) conditions (Fig. 4, right panels). The predicted values still matched the measured values (R2 = 0.963 for
Chl; R2 = 0.975 for carotenoids). Therefore, we concluded that
the color information from the bright-field microscopic images
could be useful for predicting pigment content in the cell.
Monitoring system for transition from flagellated
cells to palmelloid cells
We aimed to create a system to monitor the transition from the
motile flagellated cells (zoospores) to immotile palmelloid/cyst
cells. Flagellated cells, which were small and oval in shape (e.g.
1920
cells #1–#4 and #7 in Fig. 5A), were enriched in culture 1–2 d
after release to new medium (Wayama et al. 2013). In contrast,
palmelloid/cyst cells, which are large and spherical in shape (e.g.
cells #5 and #6 in Fig. 5A), emerged later in incubation. When
the OuterArea and OuterAxisRatio of each cell were scatter
plotted after analysis of 341 algal cells in 66 images, flagellated
cells were enriched in the left region (Fig. 5B, diamonds). To
build a classifier for the flagellated cells, we applied random
forests, a supervised ensemble machine-learning method that
achieves high classification accuracy by combining multiple decision trees constructed from a resampled input of training data
(Breiman 2001). The classifier built from the 341-cell data classified the same cells with an accuracy of 100%, i.e. all cells
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
Image-processing software for microalgae
Table 1 Description of outputs from HaematoCalMorph
Item
Description
(A) Output of single-cell data (_SingleCellData.tsv files)
Name
Name of the image file in which the cell region belongs
ID
Serial number of the cell region by image
OuterCenterX
x-coordinate of the center of a cell region
OuterCenterY
y-coordinate of the center of a cell region
Type
Cell region type
OuterArea
Area of a cell region
OuterOutlineLength
Length of outline of a cell region
OuterMaxRadius
Maximum value of the distance from the center of a cell region to the
edge of the cell region
OuterLongAxisLength
Distance between the two points at which the long axis crosses the edge of
the cell region (long axis is defined as an extended line segment of the
OuterMaxRadius)
OuterShortAxisLength
Distance between two points at which the short axis crosses the edge of
the cell region (short axis is defined as a perpendicular line segment
across the long axis)
OuterAxisRatio
Ratio of OuterLongAxisLength to OuterShortAxisLength
OuterRoundFitness
Goodness of fit of a cell region outline to an ellipse
OuterChordiogramDistance
Distance of chordiograms between a cell region and a reference region
InnerArea
Area in a cell region with lower intensity values than the threshold
(a region with colors)
InnerOutlineLength
Length of the outline of the region in a cell with colors
AreaRatio
Ratio of inner area to outer area
DistanceFromCellCenterToInnerCenterOfMass
Distance from the cell center to the center of the region with colors
AngleFromInnerCenterOfMassToFarEndOfLongAxis
Angle between the long axis and line segments from the end of the long
axis to the center of the region with colors
OuterTotalRedIntensity
Sum of intensity values of the red image in a cell region
OuterTotalGreenIntensity
Sum of intensity values of the green image in a cell region
OuterTotalBlueIntensity
Sum of intensity values of the blue image in a cell region
OuterMeanRedIntensity
Average of intensity values of the red image in a cell region
OuterMeanGreenIntensity
Average of intensity values of the green image in a cell region
OuterMeanBlueIntensity
Average of intensity values of the blue image in a cell region
InnerTotalRedIntensity
Sum of intensity values of the red image in the colored region in a cell
InnerTotalGreenIntensity
Sum of intensity values of the green image in the colored region in a cell
InnerTotalBlueIntensity
Sum of intensity values of the blue image in the colored region in a cell
InnerMeanRedIntensity
Average of intensity values of the red image in the colored region in a cell
InnerMeanGreenIntensity
Average of intensity values of the green image in the colored region in a
cell
InnerMeanBlueIntensity
Average of intensity values of the blue image in the colored region in a cell
(B) Output of the summarized data (_ImageSummary.tsv file and _Summary.tsv file)a
Name
Name of image (_ImageSummary.tsv file) or name of subfolder
(_Summary.tsv file)
Algae
Number of regions judged as algae
Colorless
Number of regions judged as colorless cells
Touch
Number of regions that touch the external margin of the image
Other
Number of regions not classified in the above three categories
CellCount
Total number of recognized regions
AlgaeOuterArea
Average of OuterArea of Algae
AlgaeOuterOutlineLength
Average of OuterOutlineLength of Algae
(continued)
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
1921
S. Ohnuki et al.
Table 1 Continued
Item
Description
AlgaeOuterMaxRadius
Average of OuterMaxRadius of Algae
AlgaeOuterLongAxisLength
Average of OuterLongAxisLength of Algae
AlgaeOuterShortAxisLength
Average of OuterShortAxisLength of Algae
AlgaeOuterAxisRatio
Average of OuterAxisLength of Algae
AlgaeOuterRoundFitness
Average of OuterRoundFitness of Algae
AlgaeOuterChordiogramDistance
Average of OuterChordiogramDistance of Algae
AlgaeInnerArea
Average of InnerArea of Algae
AlgaeInnerOutlineLength
Average of InnerOutlineLength of Algae
AlgaeAreaRatio
Average of AreaRatio of Algae
AlgaeDistanceFromCellCenterToInnerCenterOfMass
Average of DistanceFromCellCenterToInnerCenterOfMass of Algae
AlgaeAngleFromInnerCenterOfMassToFarEndOfLongAxis
Average of AngleFromInnerCenterOfMassToFarEndOfLongAxis of Algae
AlgaeOuterTotalRedIntensity
Average of OuterTotalRedIntensity of Algae
AlgaeOuterTotalGreenIntensity
Average of OuterTotalGreenIntensity of Algae
AlgaeOuterTotalBlueIntensity
Average of OuterTotalBlueIntensity of Algae
AlgaeOuterMeanRedIntensity
Average of OuterMeanRedIntensity of Algae
AlgaeOuterMeanGreenIntensity
Average of OuterMeanGreenIntensity of Algae
AlgaeOuterMeanBlueIntensity
Average of OuterMeanBlueIntensity of Algae
AlgaeInnerTotalRedIntensity
Average of InnerTotalRedIntensity of Algae
AlgaeInnerTotalGreenIntensity
Average of InnerTotalGreenIntensity of Algae
AlgaeInnerTotalBlueIntensity
Average of InnerTotalBlueIntensity of Algae
AlgaeInnerMeanRedIntensity
Average of InnerMeanRedIntensity of Algae
AlgaeInnerMeanGreenIntensity
Average of InnerMeanGreenIntensity of Algae
AlgaeInnerMeanBlueIntensity
Average of InnerMeanBlueIntensity of Algae
a
Name and description are the same in _ImageSummary.tsv and _Summary.tsv files.
classified as either flagellated or non-flagellated cells were confirmed by manual inspection (Fig. 5B, colored symbols). To test
performance, we classified 341 other different cells in the same
images with the trained classifier. When we used two parameters (OuterArea and OuterAxisRatio), the machine classified
the algal cells with an accuracy of 79% (Fig. 5D). When we used
all parameter values, the accuracy increased to 89% for test cells
(Fig. 5E). OuterTotalRedIntensity and OuterShortAxisLength
possessed high variable importance values, suggesting that
these parameters contributed well to building the classifier
for the trained cells. We also applied another supervised machine-learning algorithm, support vector machine (SVM), to
the same training and test cells and obtained similar accuracy
(Supplementary Table S2). Thus, these results indicate that
flagellated cells and palmelloid/cyst cells can be classified using
a machine-learning approach using quantitative and multivariate morphological data.
Monitoring system for dynamic changes of
H. pluvialis in culture
We aimed to create a monitoring system for the dynamics of
cells incubated under LL or LD conditions. To identify features
that varied by time, we applied principal components analysis
(PCA) on the average values of the 25 parameters at the
1922
different time points. PCA is an exploratory multivariate statistical technique for simplifying complex data sets and it has
been used for analyzing time-dependent changes in yeast
morphology data (Ohnuki et al. 2013).
The components of the first principal component (PC1),
which explained 51% of the variance (Supplementary
Fig. S1, Supplementary Table S3), changed over time under
both LD and LL conditions (Fig. 6A). The long axis lengths of
cells (OuterLongAxisLength, Fig. 6D) with the highest absolute
loading values to PC1 increased gradually in a time-dependent
manner (Fig. 6B). Likewise, the values of seven other parameters with the high absolute loading values (schematic drawings
are shown in Fig. 6D) changed gradually over time
(Supplementary Fig. S2). Accordingly, the eight parameters
presented in Fig. 6D could be used to monitor time-dependent
changes of the algal cells in culture. Among the parameters with
the low absolute loading values to PC1, InnerMeanGreenIntensity (average intensity of the green channel in the colored
region) and OuterMeanGreenIntensity (average intensity of
the green channel in the cell region) significantly decreased at
day 12 under LL conditions but showed no obvious change
under LD conditions (Fig. 6C; Supplementary Fig. S2). Both
parameters exhibited high loading values to PC3. This suggests
that PC3 could be used to distinguish between LL and LD
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
Image-processing software for microalgae
LL
2
LD
8
6
1
9
R2= 0.963
4
4
0
8
0
10
0
2
4
6
7
8
9
10 11 12 13
Cell ID
5
6
7
8
9
10 11 12 13
Cell ID
Fig. 3 Reliability and reproducibility of the data calculated by the
software. Long (left panel) and short (right panel) axes of algal cells
were measured manually by three different people (red, green and
blue crosses) and by the software (diamonds). The image shown represents the processing results of the software. Explanations of superimposed lines and letters are given in the Materials and Methods.
conditions. Because these parameters are negatively correlated
with carotenoid content, this implies accumulation of carotenoids under LL conditions. In addition, we estimated the number
of flagellated cells and immotile cells in culture using the
random forests classifier. We demonstrated that the classified
flagellated cells appeared transiently, with a peak at day 1 under
LD conditions and at day 2 under LL conditions (Fig. 6E). Thus,
we monitored both time-dependent and condition-dependent
changes in the algal cells with the quantified data.
Discussion
Our proposed image-based monitoring system is composed of
HaematoCalMorph and accompanying multivariate statistical
techniques.
The
digital
image-processing
software
HaematoCalMorph recognized H. pluvialis cells as well as colorless dead cells correctly. We easily obtained data on quantitative morphological and color information from single cells
using the software, which is in contrast to the excessive time
and labor needed to collect such data manually. Using
HaematoCalMorph and a random forests classifier, we readily
detected mobile flagellated cells and immotile palmelloid/cyst
cells. Furthermore, by performing PCA, we identified several
6
8
C a ro t e n o i d
2
4
1
0
2
4
Measured (µg/ml)
4
4
R2= 0.975
0
Short axis length (µm)
6
7
8
9
0
5
Long axis length (µm)
6
7
8
9
5
6
2
y = 1.019x
R2= 0.992
2
Estimated (µg/ml)
y = 0.992x
13
5
0
3
6
11
12
4
Chlorophyll
6
R2= 0.999
2
6
7
y = 0.922x
2
3
4
5
Estimated (µg/ml)
y = 1.000x
6
0
1
2
3
Measured (µg/ml)
Fig. 4 Estimation of the amounts of Chl and carotenoids. Measured
values (x-axes) of Chl (upper panels, green) and carotenoids (lower
panels, red) were obtained by spectrometry of the DMSO extract of
cells. To obtain predicted values (y-axes) from HaematoCalMorph
data, regression equations built from LL samples with multiple
linear regression analysis (Supplementary Table S1) were applied
to the same trained samples (left panels) and the other test LD samples (right panels). Regression equations and R2 values of the leastsquares analysis to explain predicted values from measured values are
included in the plots.
features suitable for studying cell dynamics under different
light conditions.
Parameters of HaematoCalMorph
The output parameters of HaematoCalMorph contained rich
information for characterizing the algal cells. Our software
quantifies images using 25 parameters, including cell size,
shape and pixel intensity. Of those, nine are obtained from
direct image-processing output and the other 16 are calculated
from a combination of parameters. Because no parameter is a
linear combination of the others, strong semantic correlations
with each other are not likely. For example, a high correlation
between cell area and cell outline length would be observed in
some cases, but they may be independent when the cell does
not have a coccoid form. The apparent redundancy of the parameters is not preferable, but measures from multiple parameters may improve classification of flagellated cells by a random
forests classifier.
Single-cell measurement with HaematoCalMorph
The ability to measure single cells is beneficial for studying algal
cells. First, it allows one to discriminate flagellated zoospores
from other cells. Although the zoospore is thought to be a
transient stage in the life cycle (Lee and Ding 1994,
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
1923
S. Ohnuki et al.
A
1
2
4
3
7
5
6
C
Long axis / short axis
1
1.2
1.4
training 341 cells
with 2 parameters
accuracy = 100%
4
25
100
Cell area (µm2)
4
25
100
Cell area (µm2)
E
Long axis / short axis
1
1.2
1.4
test 341 cells
with 2 parameters
accuracy = 79%
4
25
100
Cell area (µm2)
test 341 cells
with 25 parameters
accuracy = 89%
Long axis / short axis
1
1.2
1.4
D
training 341 cells
with 25 parameters
accuracy = 100%
Long axis / short axis
1
1.2
1.4
B
4
25
100
Cell area (µm2)
Fig. 5 Discrimination of flagellated cells from other cells. (A) An analyzed image of the flagellated cell-enriched sample is shown. Explanations of
superimposed lines and letters are given in the Materials and Methods. Arrows and arrowheads indicate flagellated cells and palmelloid cells,
respectively, as judged by manual inspection. (B–E) Discrimination of flagellated cells from other cells by a random forests classifier. Cell size
(area) and the ratio of long axis length to short axis length of the trained sample (B and C) and test sample (D and E) were scatter plotted.
Diamonds and circles indicate flagellated cells and palmelloid/cyst cells, respectively, determined by manual inspection. When a judgment by the
machine matched that by manual inspection, the cells are shown in blue (flagellated cells) or red (palmelloid/cyst cells). B and D, predictions with
two parameters, OuterArea and OuterAxisRatio. C and E, predictions with 25 parameters.
1924
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
Image-processing software for microalgae
5
10
140
130
120
Averaged green intensity
in the colored region (AU)
120
110
100
90
Length of long axis of
the cell region (µm)
0
150
C
110
B
130
A
15
0
5
Day
15
E
InnerTotalRedIntensity
OuterOutlineLength
OuterLongAxisLength
InnerArea
+
OuterMaxRadius
OuterArea
OuterShortAxisLength
OuterTotalRedIntensity
Fraction of flagellated cells
0
0.05
0.1
0.15
D
10
Day
0
5
10
15
Day
Fig. 6 Dynamic changes of H. pluvialis in culture. (A) Principal components analysis. Time-course plot of PC scores of PC1. (B and C) Time-course
plots of the average value of the length of the long axis (B) and the average green intensity of the colored region in the cell (C). (D) Schematic
drawing of parameters that were significantly correlated with PC1. Eight parameters with high correlation to PC1 (jloadingsj >0.95) are depicted.
Ovals and rough shapes indicate cells and the colored region in the cell, respectively. A cross in the left cell indicates the center of the cell region.
(E) Appearance of flagellated cells in culture. A time-course plot of the proportion of flagellated cells classified with the random forests classifier
to total algal cells is shown. In A–C and E, open and filled circles indicate LD and LL conditions, respectively.
Kobayashi et al. 1997b), molecular mechanisms for zoospore
formation are unknown. Statistical characterization of zoospores would help to differentiate between these cell types
and to elucidate transition mechanisms. Secondly, with our
software, the pigment content in each cell can be estimated.
Although methods for quantifying carotenoids and Chl content
have been reported (Kamath et al. 2005, Lababpour and Lee
2006, Yuan et al. 1996, Ranga et al. 2009), the estimated values
are always averaged over culture or weight. We showed in this
study that single-cell values for the pigment are useful for monitoring cell dynamics. In addition, single-cell measurements
would be advantageous when the deviation among cells is large.
Application of HaematoCalMorph to physiological
studies
Simultaneous acquisition of morphological information on a
large number of algal cells using HaematoCalMorph enables
us to monitor dynamic changes in culture. By performing
PCA, we obtained eight parameters suitable for monitoring
time-dependent changes and one parameter suitable for distinguishing different light conditions. Ultrastructural threedimensional reconstruction of algal cells revealed a marked increase in oil droplets containing astaxanthin and degeneration
of chloroplasts in the transition from palmelloid cells to cyst
cells during encystment (Wayama et al. 2013). However, that
study based on single-cell analysis would require quantitative
and statistical analysis with multiple cells. In this
study, we demonstrated a statistically significant decrease in
InnerMeanGreenIntensity at day 12 under LL conditions. The
decrease in this parameter value implied the accumulation of
astaxanthin having absorbance in the green region. The results
of this study are consistent with the ultrastructural observations of Wayama et al. (2013) and other physiological studies
(Kobayashi et al. 1992, Li et al. 2010). This implies that quantitative analyses with multiple cells using HaematoCalMorph are
useful to obtain statistically reliable conclusions about physiological changes in culture.
Quality of color, bright-field microscopic images
To obtain good results with the software, image quality is important. Well-contrasted images allow the software to detect
the edges of cells accurately. Cells should be dispersed in the
image because clustered cells are recognized as ‘other’ and
excluded from further analysis, resulting in a reduction in the
number of analyzed cells. Dustless and monotone background
images are desirable because features such as staining and dust
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
1925
S. Ohnuki et al.
could also be detected as colorless cells, resulting in an apparent
increase in dead cells.
In summary, our software obtained morphological information on single cells from multiple viewpoints simultaneously.
Multivariate statistical techniques enabled us to monitor population dynamics in a culture. Easy and immediate analysis with
our software could be used to optimize culture conditions,
assist with breeding algal strains to achieve higher carotenoid
content and/or for quality control of biotechnological processes without laborious pigment extraction steps. Because
the system is basic and general, it may be applicable to other
useful algae, including oil-producing Botryococcus, Chlorella and
Dunaliella.
Materials and Methods
Culture and observation conditions
A culture strain of H. pluvialis (K-0084) was obtained from the
Scandinavian Culture Collection of Algae and Protozoa
(SCCAP) at the University of Copenhagen. The culture
medium used was TAP medium (without agar) (Kasai et al.
2009) (http://mcc.nies.go.jp/02medium-e.html#tap). Cyst cells
were inoculated into 20 ml of TAP medium in 100 ml
Erlenmeyer flasks. Cells were incubated at 25 C under 12 h
light/12 h dark (LD) conditions or continuous light (LL) without
shaking. The light intensity was set to 100 mmol photons
m2 s1 using daylight fluorescent bulbs. At the indicated day
after inoculation, an aliquot of culture was harvested to determine the cell concentration and pigment amount and to acquire cell images. To observe flagellated cells, inoculated cells
were incubated at 25 C under continuous light (100 mmol photons m2 s1) for 1 d. For microscopic observation, 10 ml of
culture was added to a glass slide and covered with a coverslip.
Cells were observed using an Olympus BX52 microscope with
an oil immersion 40 objective lens (Olympus). Images were
obtained with a color CCD camera, DP70 (Olympus). Amounts
of Chl and carotenoids were spectrophotometrically determined by the method of Wellburn using dimethylsulfoxide
(DMSO) as a solvent for pigment extraction (Wellburn 1994).
Image-processing algorithms of the software
Step 1. Pre-processing. To reduce computing time, the size of
an original image (Fig. 1A) was reduced to half for each axis
(1,020768 pixels). The image was decomposed into three
channels (referred to as R, G and B for red, green and blue
channels, respectively) to create grayscale images. The grayscale
images were standardized so that the modes of pixel intensity
of the three channels were equal to their median by value
shifting. If the intensity values were >256 or <0, they were
set to 255 and 0, respectively. Then, levels of each standardized
image were adjusted so that the minimum and maximum
intensities were 0 and 255, respectively (referred to as leveladjusted channel images) (Fig. 1B–D). The level-adjusted
1926
channel images were converted to binary format images
(referred to as binarized channel images) (Otsu 1979).
Step 2. Edge detection and region segmentation. To detect
edges with high accuracy, we focused on two types of brightness differences. Pixels for which the local brightness changed
sharply (boundary of cells) or a boundary of global brightness
difference was observed (boundary of pigmented regions in
cells); they were identified as edges and converted to boundary
lines that segmented an image into regions. First, the edges of
the local brightness changes were detected on the level-adjusted R channel images using the Canny algorithm (Canny
1986). The detected edges were connected using the watershed
algorithm (Beucher 1982) (Fig. 1E). Secondly, the edges of the
global brightness difference were detected by tracing bright
regions of the binarized image of pigmented regions (Fig. 1F).
Because pigmented regions are chromatic, the image of the
pigment regions was obtained by calculating the Euclidean distance of each pixel brightness from the line x = y = z in threedimensional RGB space after subtraction of a mean value of
RGB for each pixel. Then, both of the detected edges were
combined after removing edges of the local brightness changes
in the bright regions of the binarized image of the pigment
regions. The edges were thinned with Hilditch’s algorithms
(Hilditch 1969) and were converted to boundary lines of 1
pixel width. The spurs generated by the Hilditch’s algorithms
were pruned by removing the end-points of the lines one after
another (Fig. 1G). These operations were done with an
embedded ImageJ toolkit (Schneider et al. 2012).
Step 3. Foreground detection. To detect cell regions, the segmented regions were classified into background regions and
foreground regions, whereby the foreground regions were
expected to include the cell regions. We combined two classification strategies. One effectively detected cell regions with
heterogeneous color intensities based on the average and variance of the pixel intensities in the regions. We used these parameters because pixel intensities were lower and more variable
in cells than in the background regions. As standards, the average and variance of pixel intensity in the temporal foreground
regions and background regions of the level-adjusted R channel
image were calculated (i.e. foreground average standard, foreground variance standard, background average standard and
background variance standard). Then, for each segmented
region, the average and variance of the pixel intensities of the
level-adjusted R channel image were calculated. The segmented
regions were classified into foregrounds and backgrounds as
follows. (i) Segmented regions with a lower average of intensities than the foreground average standard were classified as
foregrounds. Segmented regions with higher average of intensities than the background average standard were classified as
backgrounds. (ii) Among the rest of the segmented regions,
regions with higher variance of intensities than the foreground
variance standard were classified as foregrounds. Segmented
regions with a lower variance of intensities than the
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
Image-processing software for microalgae
background variance standard were classified as backgrounds.
(iii) The rest of the segmented regions were classified as foregrounds (Fig. 1H). The other strategy effectively detected colorless dead cell regions. First, R, G and B channel images were
converted to binary format images after level adjustment as
described in Step 1. Pixels in dark regions of at least one binarized channel image were judged as foreground (Fig. 1I). To
combine the results of the two strategies, pixels were classified
as foregrounds if they were judged to be a foreground region by
either method. In addition, if the number of foreground pixels
in a segmented region was larger than that of the background
pixels in the same region, the regions were classified as foregrounds. Finally, to reduce misidentification of non-cell regions,
foreground regions with <150 pixels were reclassified as backgrounds (Fig. 1J). The threshold can be configured via a graphical user interface (GUI) of the software.
Step 4. Classification of cell region. Foreground regions as
judged above were classified into cell and non-cell regions.
First, regions that touched an external margin of the image
were classified as ‘touch’ and excluded from further analysis.
Since most objects of biological origin (H. pluvialis cells in palmelloid and cyst stages and colorless dead cells) have round
shapes, resemblance to a circle would be a good measure of cell
regions. Considering future extensibility, we employed a measure chordiogram to assess resemblance to a circle since this
measure can be easily applied to other complex shapes
(Toshev et al. 2010). We calculated a chordiogram of every
foreground region and compared it with a reference chordiogram of a circle with a 50 pixel diameter. We judged the foreground regions with chordiogram distances (dissimilarity of the
two chordiograms) <0.5 as cell regions and the rest of the
foreground regions as ‘other’ regions (Fig. 1K). Among the
cell regions, colored regions were classified as ‘algae’ and the
others as ‘colorless’ (Fig. 1L).
Step 5. Calculation of morphological parameters. To describe
cells with quantitative values, we set parameters and calculated
measures for each cell region (Table 1A). Thirteen cell shape
parameters and 12 pixel intensity parameters from R, G and B
channels were calculated. The averages of the above parameter
values and the number of cells were also calculated for each
image (Table 1B).
Output of HaematoCalMorph
The output consisted of analyzed images and tab-delimited text
files (Fig. 2). A folder with an identical name to that of the input
folder contained analyzed images and a text file. The analyzed
images represented cell ID, outline of cell region, pigment regions, long and short axes, and gravity center superimposed on
the original images. Outlines of cells judged as algae (light blue),
pigmented regions in cells judged as algae (yellow), ‘touched’
cell areas that contacted the external margin of the image
(pink), colorless dead cell regions (red) and other unclassified
regions (green) are indicated. Line segments in the recognized
regions indicate long and short axes. Numerals and crosses in
the recognized regions indicate IDs in the image and gravity
centers of the regions, respectively. The text file named ‘(folder
name or subfolder name)_SingleCellData.tsv’ consisted of
information on the recognized cell regions. Each row consisted
of a different data point: image ID, cell ID, x–y coordinates of
the cell center, type of cell and the values of 25 measurements
described in Table 1A. Outside the folder, two text
files were generated. One file named ‘(input folder
name)_ImageSummary.tsv’ contained summarized information by image. The other file, named ‘(input folder
name)_Summary.tsv’, contained summarized information by
subfolder. The rows included data on image name or on the
subfolder name, number of cells by type and 25 average measurements of cells judged as algae in the image or in a subfolder
described in Table 1B.
Data analysis
The R program (http://www.R-project.org) with the standard
package was used to perform computations, the e1071 package
was used for cross-validation and parameter optimization of
SVMs (Chih-Chung and Chih-Jen 2011) and the randomForest
package was used for classification by the random forests algorithm (Breiman 2001).
Multiple linear regression analysis. To estimate the amounts
of pigment from images, a multiple regression linear model was
applied. To obtain the reference data, cells were cultured under
LL conditions and culture aliquots were harvested at 1, 2, 3, 5, 7,
9, 12 and 14 d after inoculation. We determined cell density
with a hemocytometer and the volumetric pigment concentration of DMSO-extracted samples with spectrometry, and the
acquired cell images were analyzed using HaematoCalMorph.
The determined volumetric pigment concentration was converted to cellular pigment content by dividing by cell density. A
multiple linear regression attempted to model the relationships
between explanatory variables (averaged pixel intensities in a
colored region of H. pluvialis cells) and a response variable (cellular pigment content) (Supplementary Table S1). To validate
the equations obtained from the regression analysis, we cultured cells under LD conditions and obtained the other data set
as described above.
Random forests classifier. To discriminate flagellated cells
from other cells, a random forests classifier was used
(Breiman 2001). Sixty-six images of a sample with enriched
flagellated cells were analyzed with the software, and data
from 682 cells were divided into two groups, each containing
data from 341 cells. One group was used to train the classifier to
classify flagellated cells and other cells. Two cell shape parameters (OuterArea and OuterAxisRatio) or all 25 morphological
parameters were used for training. Classifications were then
conducted according to the learned settings to classify the
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
1927
S. Ohnuki et al.
cells in the training set or the remaining 341 cells representing a
test set. Another learning algorithm, an SVM, was also used to
classify the algal cells with the same training and test sets
(Kristin and Colin 2000). Settings for the classifiers are described
in Supplementary Table S2.
Principal components analysis. PCA was performed to summarize output data of the software with scaling, and parameters that correlated significantly with the PC1 were estimated
by controlling the false discovery rate (FDR) (Storey et al. 2004).
Download site
The executable software and sample images are available at
http://sunlight.k.u-tokyo.ac.jp/HaematoCalMorph.
Supplementary data
Supplementary data are available at PCP online.
Funding
This work was supported by Core Research for Evolutional
Science and Technology (CREST), Japan Science and
Technology Agency.
Acknowledgments
We thank Kaori Takita for technical assistance, and members of
the Laboratory of Signal Transduction and the Laboratory of
Plant Life System, GSFS, the University of Tokyo, for discussions.
S.O. was a Research Fellow of the Japan Society for the
Promotion of Science.
Disclosures
The authors have no conflicts of interest to declare.
References
Beucher, S. (1982) Watersheds of functions and picture segmentation.
In Acoustics, Speech, and Signal Processing, IEEE International
Conference on ICASSP ’82. pp. 1928–1931.
Boussiba, S. (2000) Carotenogenesis in the green alga Haematococcus
pluvialis: cellular physiology and stress response. Physiol. Plant. 108:
111–117.
Boussiba, S., Bing, W., Yuan, J.P., Zarka, A. and Chen, F. (1999) Changes
in pigments profile in the green alga Haeamtococcus pluvialis
exposed to environmental stresses. Biotechnol. Lett. 21: 601–604.
Breiman, L. (2001) Random forests. Machine Learning 45: 5–32.
Canny, J. (1986) A computational approach to edge-detection. IEEE
Trans. Pattern Anal. Machine Intell. 8: 679–698.
Chih-Chung, C. and Chih-Jen, L. (2011) LIBSVM: a library for support
vector machines. ACM Trans. Intell. Syst. Technol. 2: 1–27.
1928
Collins, A.M., Jones, H.D.T., Han, D.X., Hu, Q., Beechem, T.E. and
Timlin, J.A. (2011) Carotenoid distribution in living cells of
Haematococcus pluvialis (Chlorophyceae). PLoS One 6: e24302.
Hilditch, C.J. (1969) Linear skeletons from square cupboards. In
Machine Intelligence. Edited by Meltzer, B. and Michie, D. pp.
403–420. Edinburgh University Press, Edinburgh, UK.
Hoffman, Y., Aflalo, C., Zarka, A., Gutman, J., James, T.Y. and
Boussiba, S. (2008) Isolation and characterization of a novel chytrid
species (phylum Blastocladiomycota), parasitic on the green alga
Haematococcus. Mycol. Res. 112: 70–81.
Kaczor, A. and Baranska, M. (2011) Structural changes of carotenoid
astaxanthin in a single algal cell monitored in situ by Raman spectroscopy. Anal. Chem. 83: 7763–7770.
Kaczor, A., Turnau, K. and Baranska, M. (2011) In situ Raman
imaging of astaxanthin in a single microalgal cell. Analyst 136:
1109–1112.
Kamath, S.B., Chidambar, S., Brinda, B.R., Kumar, M.A., Sarada, R. and
Ravishankar, G.A. (2005) Digital image processing—an alternate
tool for monitoring of pigment levels in cultured cells with special
reference to green alga Haematococcus pluvialis. Biosens.
Bioelectron. 21: 768–773.
Kasai, F., Kawachi, M., Erata, M., Mori, F., Yumoto, K., Sato, M. et al.
(2009) NIES-collection list of strains, 8th edition. Jpn. J. Phycol.
(Sôrui) 57 (Suppl.): 1–350.
Kobayashi, M., Kakizono, T. and Nagai, S. (1991) Astaxanthin production by a green alga, Haematococcus pluvialis accompanied with
morphological changes in acetate media. J. Ferment. Bioeng. 71:
335–339.
Kobayashi, M., Kakizono, T., Nishio, N. and Nagai, S. (1992) Effects of
light-intensity, light quality, and illumination cycle on astaxanthin
formation in a green-alga, Haematococcus pluvialis. J. Ferment.
Bioeng. 74: 61–63.
Kobayashi, M., Kakizono, T., Nishio, N., Nagai, S., Kurimura, Y. and
Tsuji, Y. (1997a) Antioxidant role of astaxanthin in the green alga
Haematococcus pluvialis. Appl. Microbiol. Biotechnol. 48: 351–356.
Kobayashi, M., Katsuragi, T. and Tani, Y. (2001) Enlarged and astaxanthin-accumulating cyst cells of the green alga Haematococcus
pluvialis. J. Biosci. Bioeng. 92: 565–568.
Kobayashi, M., Kurimura, Y., Kakizono, T., Nishio, N. and Tsuji, Y.
(1997b) Morphological changes in the life cycle of the green alga
Haematococcus pluvialis. J. Ferment. Bioeng. 84: 94–97.
Kristin, P.B. and Colin, C. (2000) Support vector machines: hype or
hallelujah? SIGKDD Explor. Newsl. 2: 1–13.
Kurashige, M., Okimasu, E., Inoue, M. and Utsumi, K. (1990) Inhibition
of oxidative injury of biological-membranes by astaxanthin. Physiol.
Chem. Phys. Med. NMR 22: 27–38.
Lababpour, A. and Lee, C.G. (2006) Simultaneous measurement of
chlorophyll and astaxanthin in Haematococcus pluvialis cells by
first-order derivative ultraviolet–visible spectrophotometry. J.
Biosci. Bioeng. 101: 104–110.
Lee, Y.K. and Ding, S.Y. (1994) Cell-cycle and accumulation of astaxanthin in Haematococcus lacustris (Chlorophyta). J. Phycol. 30:
445–449.
Li, Y., Sommerfeld, M., Chen, F. and Hu, Q. (2010) Effect of photon flux
densities on regulation of carotenogenesis and cell viability of
Haematococcus pluvialis (Chlorophyceae). J. Appl. Phycol. 22:
253–263.
Mosleh, M.A., Manssor, H., Malek, S., Milow, P. and Salleh, A. (2012) A
preliminary study on automated freshwater algae recognition and
classification system. BMC Bioinformatics 13 (Suppl 17): S25.
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
Image-processing software for microalgae
Naguib, Y. (2000) Antioxidant activities of astaxanthin and related
carotenoids. J. Agric. Food Chem. 48: 1150–1154.
Ohnuki, S., Enomoto, K., Yoshimoto, H. and Ohya, Y. (2013) Dynamic
changes in brewing yeast cells in culture revealed by statistical analyses of yeast morphological data. J. Biosci. Bioeng. (in press, doi:
10.1016/j.jbiosc.2013.08.005).
Ohnuki, S., Kobayashi, T., Ogawa, H., Kozone, I., Ueda, J., Takagi, M.
et al. (2012) Analysis of the biological activity of a novel 24-membered macrolide JBIR-19 in Saccharomyces cerevisiae by the morphological imaging program CalMorph. FEMS Yeast Res. 12:
293–304.
Ohya, Y., Sese, J., Yukawa, M., Sano, F., Nakatani, Y., Saito, T.L. et al.
(2005) High-dimensional and large-scale phenotyping of yeast mutants. Proc. Natl Acad. Sci. USA 102: 19015–19020.
Otsu, N. (1979) A threshold selection method from gray-level histograms. IEEE Trans. Syst, Man Cybernet. 9: 62–66.
Palozza, P. and Krinsky, N. (1992) Astaxanthin and canthaxanthin are
potent antioxidants in a membrane model. Arch. Biochem. Biophys.
297: 291–295.
Ranga, R., Sarada, A.R., Baskaran, V. and Ravishankar, G.A. (2009)
Identification of carotenoids from green alga Haematococcus pluvialis by HPLC and LC-MS (APCI) and their antioxidant properties. J.
Microbiol. Biotechnol. 19: 1333–1341.
Rodenacker, K., Hense, B., Jütting, U. and Gais, P. (2006) Automatic
analysis of aqueous specimens for phytoplankton structure
recognition and population estimation. Microsc. Res. Tech. 69:
708–720.
Schneider, C., Rasband, W. and Eliceiri, K. (2012) NIH Image to ImageJ:
25 years of image analysis. Nat. Methods 9: 671–675.
Storey, J., Taylor, J. and Siegmund, D. (2004) Strong control, conservative point estimation and simultaneous conservative consistency of
false discovery rates: a unified approach. J. R. Stat. Soc. B 66:
187–205.
Toshev, A., Taskar, B. and Daniilidis, K. (2010) Object detection via
boundary structure segmentation. 2010 IEEE Conference on
Computer Vision and Pattern Recognition. pp. 950–957.
Walker, R.F., Ishikawa, K. and Kumagai, M. (2002) Fluorescenceassisted image analysis of freshwater microalgae. J. Microbiol.
Methods 51: 149–162.
Wayama, M., Ota, S., Matsuura, H., Nango, N., Hirata, A. and
Kawano, S. (2013) Three-dimensional ultrastructural study of oil
and astaxanthin accumulation during encystment in the green
alga Haematococcus pluvialis. PLoS One 8: e53618.
Wellburn, A. (1994) The spectral determination of chlorophyll a and
chlorophyll b, as well as total carotenoids, using various solvents
with spectrophotometers of different resolution. J. Plant Physiol.
144: 307–313.
Yuan, J.P., Gong, X.D. and Chen, F. (1996) Separation and identification
of astaxanthin esters and chlorophylls in Haematococcus lacustris
by HPLC. Biotechnol. Tech. 10: 655–660.
Plant Cell Physiol. 54(11): 1917–1929 (2013) doi:10.1093/pcp/pct126 ! The Author 2013.
1929