Morphological Characterization of the

Dimorphic Yeast KIuyveromyces

marxianus var. marxianus NRRLy2415 by
Semi-Automated Image Analysis
D. G. OShea and P. K. Walsh*
Biochemical Engineering Research Group, School of Biological Sciences,
Dublin City University, Dublin 9, Ireland
Received December 18, 1995/Accepted March 19, 1996

A semiautomatic image analysis method has been devel- Qualitative characterization of morphology is com-
oped to characterize the morphology of the dimorphic monly achievzd by visual inspection of the fermentation
yeast Khyveromyces marxianus var. marxianus (for- broth. When a quantitative analysis is required, it is
merly fragilis) NRRLy2415 undergoing alcoholic fermen-
tation of cheese whey permeate. The method is capable usually necessary to resort to an image processing sys-
of separating cells into six defined categories, varying tem. In recent years, much work has been reported on
from simple ovoid yeast cells to branched mycelial cells. the use of image analysis to describe the morphological
A sample size of 300 cells was found to be sufficient to character of fermentation broths containing a variety
obtain a statistically significant categorization. The pro- of cell types: actinomycetes (Reichl et al., 1992; Warren
cessing time for a sample was found to be approximately
90 min. In addition to qualitative characterization, the et al., 1995), yeast (Huls et al., 1992; Pons et al., 1993),
method permits the measurement of geometric proper- mammalian (hybridoma) cells (Tucker et al., 1994),
ties such as the width, length, and volume of individual plant cells (Kieran et al., 1993), and fungi in pelleted
cells or clusters of cells. When the cells analyzed by the (Cox and Thomas, 1992) and filamentous forms (Packer
automatic method were categorized on a manual basis, et al., 1992; Tucker et al., 1992).
the error level in the automatic routine was found to be
less than 3%. 0 1996 John Wiley & Sons, Inc. The complexity and diversity of biological systems
Key words: yeast dimorphism morphology image requires that the parameters required for morphological
ana Iysis characterization will vary between cell types (bacterial,
yeast, fungal, mammalian, or plant cells) and within the
cell type (pelleted or filamentous forms). For example,
INTRODUCTION
a yeast cell can usually be described by measuring the
The performance of a bioreactor is greatly influenced length of the major and minor axis and calculating pa-
by the morphological character of the cells in the broth. rameters such as cell volume, by assuming that the shape
The rheological properties of the broth are determined of an individual yeast cell is that of a prolate spheroid
by such factors as the concentration of biomass and (Pons et al., 1993).With filamentous cells, typically mea-
its morphology. Moreover, morphology affects oxygen sured and calculated parameters include the length of
transfer, nutrient consumption, and mixing. Process the main hypha, total hyphal length, branching fre-
variables influencing morphology include shear stress, quency, hyphal growth unit (Thomas, 1992), and cell
type of inoculum, medium constituents, pH, and dis- volume (Packer et al., 1992). For pelleted cells, distinc-
solved oxygen concentration. The morphological char- tion can be made between “hairy” and “smooth” pellets
acter of cells may, in some cases, influence the outcome (Cox and Thomas, 1992).
of microbial reactions. The addition of n-hexadecane to In the majority of the cultures described above, there
a culture of Penicillium chrysogenum at the beginning is relatively little change in the morphological character
of a fermentation has been shown to result in a reduction of the cells over the course of a fermentation. However,
in penicillin production due to a change in the cell mor- there are many occurrences, in both the fungal and
phology from smooth compact pellets to fluffy loose bacterial kingdoms, of cells which exhibit dimorphism;
pellets (Peng and Chen, 1994). Increased yields of bio- the phenomenon whereby the mycelial habit of growth,
mass and gibberellic acid have been related to differ- with cells in hyphal or filamentous form, is transformed
ences in morphological character of Fusarium monili- by some change in cultural or environmental conditions,
forme grown in fermenters with dissimilar mixing so that a yeastlike or unicellular morphology is adopted
geometries (Priede et al., 1995). at the cellular level (San-Blas and San-Blas, 1984). Di-
morphism is displayed in a myriad of clinically and in-
* To whom all correspondence should be addressed. Telephone: dustrially important yeast and fungi. The pullulan pro-
353 1 7045393; fax: 353 1 7045412; e-mail: [email protected] ducer, Aureobasidium pullulans, has been widely

Biotechnology and Bioengineering, Vol. 51, Pp. 679-690 (1996)

0 1996 John Wiley & Sons, Inc. CCC 0006-3592/96l060679-12
studied (McNeil et al., 1989; Reeslev and Jensen, 1995; flask culture which had previously been inoculated from
Yamasaki et al., 1993), but even the common yeast Sac- an agar plate. A flask was withdrawn at each sample
charomyces cerevisiae has been shown to exhibit dimor- point for analysis.
phic behaviour when grown in continuous culture (Hill
and Robinson, 1988).
Analytical Methods
Despite the obvious importance of dimorphic strains,
the use of image analysis to quantitatively characterize Biomass concentrations were estimated by filtering 50-
morphological transitions has been scant. The present 100-mL samples of broth through Whatman No. 1filter
work is concerned with the behavior, in batch shake- papers (Whatman Inc., Maidstone, England), washing
flask culture, of Kluyveromyces marxianus var. marx- the resultant filter cakes with 100mL of deionized water,
ianus NRRLy2415, a strain which has previously been and then drying the filter at 105°C for 24 h to obtain
shown to be dimorphic in character (Walker and the biomass dry cell weight. Lactose was determined
O'Neill, 1990). K. marxianus is extensively used for the using the dinitrosalicylic acid method (Miller, 1959).
fermentation of cheese whey permeate to ethanol (Vi- Ethanol concentration was determined using a Carla
enne and von Stockar, 1985) and also possesses a wealth Erba HRGC S300 Mega Series gas chromatograph
of hydrolytic enzymes, including galactosidase (Ku and (Carla Erba Strumentazione, Milan, Italy) with flame
Hang, 1992), inulinase (Rouwenhorst et al., 1990), and ionisation detection. The column was packed with 10%
polygalacturonases (Schwan and Rose, 1994). carbowax 20M on chromosorb WAW 80/100 mesh.
A semiautomatic image analysis routine is described Measurements of morphological parameters were
which permits the quantitative characterization of K. made using a Leica Q5WMC image processing and anal-
marxianus populations displaying morphologies varying ysis system (Leica Cambridge, Cambridge, England),
from simple ovoid yeast cells to branched mycelial using a Sony XC-75CE CCD monochrome video cam-
forms. This routine is shown to be flexible and capable era (Sony Corporation, Japan) linked to an Olympus
of handling the wide range of morphologies that are BX40 brightfield microscope (Olympus Optical Co., Ja-
displayed over the course of a batch fermentation pro- pan) using 200X magnification. From the video camera,
ducing ethanol from cheese whey permeate. In addition, a digitized image (of 720 X 512 square pixels) is captured
the routine permits the calculation of important geomet- with gray scale values ranging from 0 to 255. The
ric properties of cells, such as effective length and Q500MC system contains an Intel 486 DX2 (66 MHz)
volume. processor. Sample preparation involved suitably dilut-
ing the broth with distilled water, placing 20 p L on a
glass slide, and covering with a cover slip.
MATERIALS AND METHODS

MORPHOLOGICAL CHARACTERIZATION
Microorganism and Culture Conditions
Traditionally, dimorphic organisms have been classified
Kluyveromyces marxianus var. marxianus NRRLy2415 into two qualitative forms, one termed yeastlike (Y)
(Northern Regional Research Laboratories, Peoria, IL) and the other termed either filamentous (F) or mycelial
was the strain used in this work. The organism was (M). Populations are generally characterized in terms
maintained on YEPD-agar plates containing the follow- of %Y and %F (or %M) forms (Reeslev and Jensen,
ing composition (g L-l): technical agar (Oxoid No. 3, 1995; Walker and O'Neill, 1990). Although there is
Basingstoke, England), 30; yeast extract (Oxoid), 10; agreement that the term yeastlike describes oval to ellip-
bacteriological peptone (Oxoid), 20; and dextrose soidal cells, the filamentous form often includes elon-
(Riedel-de Haen, Seeize, Germany), 20. The medium gated cells, which are unicellular with elongation of the
used for growth in liquid culture was prepared by resus- cell's major axis, and pseudomycelia, which are elon-
pending 100g commerical whey powder L-' (Avonmore gated cells where each generation of buds remains
plc, Kilkenny, Ireland) in deionized water, removing attached to the parent cells (Walker and O'Neill, 1990).
excess protein in a Romicon PM30 hollow fiber ultrafil- Pons et al. (1993) subcategorized yeast populations in
tration unit (Romicon Inc., Woburn, MA), and suitably terms of the frequencies of single cells, double cells, and
diluting the resulting permeate to yield a medium lac- clusters of cells.
tose concentration of 50 g L-'. The medium was then The K. marxianus strain used in this work can display
supplemented with 5 g ammonium sulphate L-' (Riedel- morphologies that vary from simple ovoid yeast cells
de Haen) and the pH was adjusted to 4.5 using concen- to very long, branched (both primary and secondary
trated H2S04.The organism was cultivated on an orbital branching) almost true mycelium with up to 25 cell
shaker (133 rpm) at 30"C, using 100 mL of medium in units. To describe adequately the heterogeneity of the
250-mL, foil-covered Erlenmeyer flasks. For the fer- culture, we believe that a quantitative measurement of
mentation run described, eight such flasks were inocu- the frequency distribution of cell volume is required.
lated with 2-mL aliquots taken from a 24-h Erlenmeyer Furthermore, to characterize quantitatively the organ-

680 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 51, NO. 6, SEPTEMBER 20, 1996
 ism in the present work, it is felt that the system of
classification of cells as being either yeastlike or fila-
mentous is inadequate and that a classification into six
subgroups is required, as presented in Figure 1 and
Table I.

Image Analysis Method

The image analysis routine was developed to perform
the following functions on a cell population:
0 Separation of cells into the six classes described in
Table I.
0 Estimation of the volume of individual cells and dou-
ble cells (see Table I).
0 Measurement of the morphological properties of my-
celial cells, such as the total length, mean width, vol-
ume, and extent of branching (see Table I).
The image analysis procedure is divided into seven
stages and is depicted in Figure 2:
1. Gray image acquisition, binary thresholding, and
editing initially involve acquiring and storing enough
gray image fields (Image A, Fig. 3a) to yield an ade-
quately large sample size (300 cells would typically re-
quire six fields of view). As the microscope is not
equipped with automatic slide motion, auto-focus, or
brightness control, it is necessary to set manually a bi-
nary threshold to detect the objects of interest. In prac-
tice, the resultant binary images (Image B, Fig. 3b)
are not of sufficient quality for measurements to be
performed. Subsequent binary image processing in-
volves a number of functions: (1)Very small particulate
matter is removed by a primary area filter (whereby
objects with an area <4 pm2 are rejected); (2) budded
Figure 1. Morphological forms displayed by K. marxianus var. marx- cells which have been poorly detected are rebuilt by
ianus NRRLy2415. ( a ) Yeast, (b) elongated yeast, (c) double yeast, performing a series of dilations and erosions; (3) holes
(d) filament, (e) double filament, (f) mycelium. Bar indicates 10 ~ m . in the central parts of cells are filled; and (4) after the
preceding steps have been completed, a secondary area
filter removes all objects with an area <8 pm2. It was

Table I. Morphologies displayed by K. marxianus var. marxianus NRRLy2415: Classification and volume estimation.

Class Description Volume

Yeast Spherical or ellipsoidal single cells LW%l6

Elongated yeast Elongated ellipsoidal single cells LW%6
Filament Rodlike cell with no visible constriction LW7i-14"
Double yeast Budding yeast or budding elongated yeast containing a visible constriction at the ( L l w + L2w)7i-16
mother-bud junction
Double filament Joined filaments formed either by cell growth or mycelial fragmentation ( L ~ +wL~~ W : ) ~ ~or- W
~~n7i-14~
Mycelium Three of more cells joined together (usually composed of filaments); may be Lw7i-14d
branched or unbranched

L and W are the feret length and width of a cell, respectively. The subscripts refer to the first and second subunits of a double cell.
"The length L of a filament is the feret length when it has been measured from Image G (Fig. 2) and is the skeletonized length (where the
cell has been reduced to single pixel width and length is measured as perimetern) when it has been measured from Image H (Fig. 2).
bVolume of double filament which has been successfully segmented in Image F (Fig. 2).
'Volume of double filament which could not be segmented in Image F (Fig. 2) and has been transferred to and measured in Image H (Fig. 2).
d L is the skeletonized length of a mycelium. W is the mean width of the cell and is equal to aredlength.

O'SHEA AND WALSH: IMAGE ANALYSIS OF DIMORPHIC YEAST 68 1

found that applying this filter at stage (1) eliminated a Due to the heterogeneous nature of the cell population
number of budded cells which had been poorly detected. described within this work, no simple filter was capable
Occasionally, on the completion of step (4),an artifact of efficiently distinguishing such artifacts from cells and
larger than 8 pm2 was clearly visible on the screen. the artifact was deleted by a manual override facility.

682 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 51, NO. 6, SEPTEMBER 20, 1996
 Table II. Filters for preliminary classification of cells.
Filter
Class number Filter parameters

Single cell 1 8 < area < 15 pm2

Single cell 2 15 5 area < 25 pm2 and perimeter/
convex perimeter < 1.12
Mycelium 3 area > 60 pm2

The final binary images consist entirely of cells to be

measured and classified (Image C, Fig. 3c).
2. Preliminary cell classification involves separating
single cells and mycelia from Image C and placing them
in separate image stores. Initially, a primary classifica-
tion of single cells is performed on Image C, using two
sizehhape filters (Filters 1 and 2, as detailed in Table
11), giving rise to a preliminary single cells image (Image
D, Fig. 2). Mycelial cells are then removed from Image
C using a size filter (Filter 3 in Table II), resulting in a
preliminary mycelial cells image (Image E, Fig. 2). Once
the single cells and mycelial cells have been removed
from Image C, the remaining cells are termed “possible”
double cells (Image F, Fig. 2).
3. Primary segmentation. Image F is termed “possi-
ble” double cells because, in addition to “true” double
cells (both double yeast and double filaments), this im-
age also contains cells that are, in fact, either single cells
or mycelia, but that are not captured by the software
filters described in Table 11. To classify these cells cor-
rectly, it is necessary to perform a “segmentation” on
each cell in Image F. The segment function is a system-
inbuilt binary morphological algorithm that is used to
automatically separate touching features by reconstruc-
tion from ultimate erosion markers. Primary segmenta-
* tion proceeds by systematically (and automatically) re-
moving the largest object in Image F and placing it in
a separate image plane. A segmentation is then per-
formed on this object and the number of entities in the
frame is counted. (This primary segmentation operation
is hereafter termed Segment I . ) If the cell separates into
two subunits, the geometric properties of the subunits
are measured and the cell is classified as a double yeast
or a double filament according to the criteria of Table
111. The cell is then removed from Image F.
4. Secondary segmentation. There are a number of
cell geometries in Image F where the application of the
segment function either fails to separate double cells or
where more than two subunits arise. Extended image
processing must be performed on these cells if they are
to be classified correctly and automatically. The use of a

1,
I Figure 3. Preliminary image processing stages. (a) Image A: (portion
of) initial gray image. (b) Image B: unedited binary image. (c) Image
C edited binary image. For clarity, only a portion (approximately
one third) of the full image frame is displayed. Bar indicates 10 pm.

O‘SHEA AND WALSH: IMAGE ANALYSIS OF DIMORPHIC YEAST 683

 i
Table 111. Classification of single and double cells.

-
Cell type Classification

Yeast 1 < LIW < 2.5

Elongated yeast 2.5 5 LIW < 4
Filament LIW 2 4
Double yeast LllWl and LzlW2 < 4 Binary Outline Binary - Outline Rebuilt and Segmented
Double filament L1lWl or L2lW2 2 4
(a)

system-inbuilt binary “outline” morphological operator

proved extremely useful in correctly classifying (and
separating, if necessary) the aforementioned cells. Any
cell in Image F which failed to segment into two subunits
CI
after the application of Segment 1 was subjected to a
series of operations (hereafter termed Segments 2, 3,
and 4 ) , which were applied in sequence and which are Binary Binary - Outline Segment Operator Rebuilt and Segmented

-
described in Figure 4. (b)

Segment2 This operation proved useful for separating

double cells that segmented into more than two
subunits when primary segmentation (Segment I ) H
was performed. The binary outline operator is ap-
plied to the cell. In Figure .4a, this results in the
outline continuing through the cell-cell interface.
Using logical arithmetic, the cell is seen to separate
into two subunits. Using a number of logical and
Binary Binary - Outline Segment Operator Rebuilt without bud
rebuild functions, it is possible to recreate the origi-
nal cell image, but whereby the cells are separated (c)
at the interface, as depicted in Figure 4a. The geo-
Figure 4. Morphological operations for secondary image processing:
metric parameters of the subunits are measured and (a) Segment 2. (b) Segment 3. (c) Segment 4. Bar indicates 10 prn.
their volume calculated as an ellipsoid or cylinder as
required. The cell is then removed from Image F.
Segment3 Some double cells do not possess the neces- ever, if the difference in area between the original
sary degree of concavity at the cell-cell junction to cell and the segmented cell is 3 pixels or less, the
separate into two subunits by primary segmentation cell is deemed to be a (budding) single cell. It is
(Segment I ) and also fail to separate when the series then removed from Image F and added to the final
of outlining and logical operations in Segment 2 is single-cells image (Image G, Fig. 2).
applied (as depicted in Fig. 4b). If, however, the
segment operator is applied to the cell structure in After these steps have been completed, all the “true”
place at the end of Segment 2, and this successfully double cells in the original “possible” double cell image
separates the cells into two subunits, the unit may (Image F) have been classified (and their geometric
be rebuilt, measured, and eventually removed from properties have also been measured). If, after applying
Image F. This step is termed Segment3. If, following Segments I, 2, 3 and 4, the original cell is still intact, it
the application of the segment operator, three or is added to the original single cells image, resulting in
more subunits are present, the cell is defined as a a final single cells image (Image G, Fig. 2). If, after
mycelium and is copied to Image H (the final myce- applying Segments I , 2 , 3 and 4, the number of subunits
lial image) and simultaneously removed from Im- is greater than two, the cell is added to the original
age F. mycelial cell image, resulting in a final mycelial cells
Segment 4 If, following the application of Segments image (Image H, Fig. 2).
I , 2, and 3, the cell still fails to separate (see Fig. 5. ClassiJcationand measurement ofproperties of sin-
4c) but whereby there is a significant difference gle cells. The individual cells in Image G are subclassified
(>3 pixels) between the measured areas of the orig- as yeast, elongated yeast, or filaments according to the
inal cell and the image remaining after the applica- criteria listed in Table 111. As all cells in this image are
tion of Segments I , 2, and 3, the cell can be rebuilt, single cells and, hence, are discrete objects, measure-
measured, and removed from Image F. (This se- ment of geometric properties can be performed on the
quence of operations is termed Segment 4 . ) How- entire image in a single pass.

684 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 51, NO. 6, SEPTEMBER 20, 1996
Table IV. Classification of filaments/unbranched mycelia (in Im- 4 60
age H).
I

i-\ f. lo
Class Filter

Filament Ls18pm
Double filament 18 < L 5 36 pm 3
Unbranched mycelium L > 36pm

6. Skeletonization and measurement of geometric

properties of rnycelial cells. The algorithm for skeletoniz-
ing a mycelium and pruning the branches was developed
in a manner similar to that described by other workers / K I
(Tucker er al., 1992).The process begins by skeletonizing
the largest object in Image H. After skeletonization, the
number of tips is measured. If this equals 2, the cell is
obviously unbranched and pruning is not required. If
the number of tips is three or more, the skeletonized 0 3 6 9 12 15 18 21 24
image is pruned and the length measured. Following Time (h)
these operations, the original object is removed from
Image H. This image also contains some single filaments Figure 5. Batch growth profiles for Kluyveromyces marxianus var.
(which were not removed by the Table I1 filters) and marxianus NRRLy2415 shake-flask culture (0)biomass, (0)lactose,
(m) ethanol.
double filaments, which segmented into more than two
subunits even after applying the primary and secondary
segmentation operations (Segments I, 2,3, and 4). Such that although the mycelia only represent 11%of the
cells were classified using the criteria in Table IV. The total cells by number, they constitute one quarter of the
individual sizes of the two subunits of the double fila- entire biomass by volume.
ments cannot be measured, but the volume of the entire The classification of cells at three-hour intervals over
unit is calculated as described in Table I and the unit the course of the fermentation is displayed in Figure 6.
is categorized using the criteria of Table IV. From 0 to 6 h, the percentage of single cells decreases
7. Sorting and presentation of data. The measured and initially and the number of double yeast cells increases,
calculated morphological data from single cells, double indicating biomass growth. In the period from 6 to 9 h,
cells, and mycelia were sorted into the desired order, and the percentage of double filaments increases dramati-
the resulting graphs were produced using the Sigmaplot cally. The development of many of these cells into myce-
graphics package (Jandel Scientific, Erkrath, Germany). lia is seen to occur between 9 and 12 h. As the medium
is exhausted during the latter stages of the run, an in-
crease is noted in the percentage of single cells at the
RESULTS expense of other cells. As the inoculum was taken from
The time course of a fermentation run is presented an overnight culture, the expected similarityin the distri-
in Figure 5. A low rotation rate was chosen, as our bution of cells at the beginning and toward the end of
experience had shown that a wide variety of cell types the fermentation can also be seen.
would be present over the course of such a fermentation, The number of cells that must be analyzed is a com-
ensuring that the robustness and flexibility of the algo- promise between the accuracy desired, the total time
rithm would be sufficiently challenged. The fermen- taken to analyze the sample, and the amount of storage
tative nature of the run is evidenced by the preferential available for the images. The effect of sample size on
conversion of lactose into ethanol rather than cell mass. the measured percentage of cells in each morphological
The image analysis algorithm permits cells to be cate-
gorized in two ways: (1) by the number of cells in each Table V. Comparison of numerical classification with volumetric
category, and (2) by the percentage of the total cell classification of cells (t = 12 h).
volume occupied by cells in each category. Since, for
example, a single mycelium occupies a much greater Class By object count By volume
volume than a single yeast cell, data are normally repre- Yeast 22% 15%
sented as percentage volume. It is presumed that the Elongated yeast 26% 14%
volume of the cells is also more directly comparable to Double yeast 26% 28%
the biomass concentration. The difference between the Filament 7% 6%
Double filament 8% 12%
two classifications is presented in Table V for the 12-h
Mycelium 11% 25%
sample (in which 302 cells were analyzed). It is evident

O'SHEA AND WALSH: IMAGE ANALYSIS OF DIMORPHIC YEAST 685

 EBl Yeast Double yeast Double Filaments
Elongated yeast 6Filaments Mycelium

60 60
E
P
-m
.-0
m 40 40
-
0
0
c
$
E 20 20
r:
0
m
a,
Y-
0
s 0 O
Oh 3h 6h
E 60 60

40 40

20 20

0 0

9h 12 h 15 h
60 60
E
P
-
m0
.-
0)
-
0 40 40
0
c
E?
i2
r 20 20
0
m
a,
u-
0
s 0 0

18 h 21 h 24 h
Figure 6. Distribution of cell morphologies during the fermentation run.

category is presented in Table VI for the 12-h sample. Table VI. Effect of sample size on the distribution within each cell
To avoid any bias, the data for the sample was random- category (based on cell number, t = 12 h).
ized each time before the first 100, 200, etc. cells were Number of cells taken at random from a
categorized. A sample size of 300 cells (approximately sample of 420 cells
6 fields at an overall magnification of 200X) was deter-
mined to be sufficient in obtaining an accurate classifi- Class 100 200 300 400
cation, at a significance level of 5%. Yeast 26% 25% 22% 21%
Due to the structure of the algorithm, some cells are Elongated yeast 18% 23% 23% 23%
categorized and measured quickly although others are Double yeast 22% 22% 24% 26%
subjected to a large number of processing operations. Filament 8% 9% 9% 9%
Double filament 8% 9% 8% 9%
The effectiveness and limitations of the various compo- Mycelium 16% 12% 14% 12%
nents of the automatic image analysis routine can best

686 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 51, NO. 6, SEPTEMBER 20, 1996
Table VII. Time taken to perform the individual stages of the image analysis algorithm, (t = 12 h).
~

Stage Images Unit time Number of cells Number of fields Total time

Filter 1 (Table 11) C+D 5 slfield 30 s

Filter 2 (Table 11) C+D 5 slfield 30 s
Filter 3 (Table 11) C-+E 5 slfield 30 s
Primary segmentation (Segment I ) F 14 slcell 812 s
Secondary segmentation (Segment 2 ) F 37 Slcell 666 s
Cells sent to mycelium image F+H 37 slcell 999 s
Secondary segmentation (Segment 3) F 49 slcell 49 s
Secondary segmentation (Segment 4 ) F 49 slcell 637 s
Cells sent to single cell image F+G 45 slcell 720 s
Single cell measurement G 5 slfield 30 s
Branched mycelia H 41 slcell 410 s
Unbranched mycelia (Table VI) H 20 slcell 440 s
Filaments in mycelia image (Table VI) H 20 slcell 220 s
Double filaments in mycelia image (Table VI) H 20 skell 240 s

Total 89 min

*Numbers within parentheses indicate cells being transferred between images. Numbers without parentheses indicate the stage at which the
geometric properties of cell are measured.

be understood by observing the processing of a specific In fact, in classification terms, there are only 8 differ-
cell sample through the algorithm. This is demonstrated ences between the automatic and manual categories
for the 12-h sample in Table VII. This sample was cho- (<3%).
sen as the proportion (by cell number) of cells in any The time course of the distribution of the volumes
category is no less than 7% and no greater than 26%. and lengths of cells is depicted in Figure 7. It can be
Table VII also illustrates the time taken to perform seen that there is a both a significant broadening of and
each section of the algorithm. A comparison of the rela- a modal increase in the distribution of cell volumes in
tive time taken to perform an operation is probably the first 3 h, followed by a narrowing of and a modal
more useful than the absolute time, as this is system decrease in the distribution over the remainder of the
dependent and would be considerably quicker on a more fermentation run. The significant percentage of mycelia
powerful Quantimet (or other) image analyzer. measured in the 12-h sample of Figure 6 is seen to
To check the accuracy of the automatic image analysis correspond in Figure 7a to a large number of cells whose
routine, the cells were also processed through the algo- volumes range from 100 to 250 pm3. In Figure 7b, the
rithm in such a manner that the operator was able to distribution of cell length is seen to broaden significantly
intervene manually and inspect the correctness of the during the growth phase, reflecting the large numbers
cell classification technique. For the 12-h sample, a total of mycelia present at this stage. The length distribution
of 17 of the 302 cells (5.6%) were seen to have been is seen to narrow toward the latter stages of the fermen-
classified incorrectly, due to errors in segmentation. The tation run, with the distribution of the 24-h sample
nature of the errors are detailed in Table VIII. It can closely resembling that of the initial sample.
be seen, however, that two of the mistakes occurred
where double cells segmented at the wrong location DISCUSSION
but where no error in categorization took place and, in The bound of the filters in Tables I1 and IV were neces-
addition, a number of the other errors are self-canceling. sarily determined by trial and error. The primary func-

Table VIII. Classification of segmentation errors made by algorithm (t = 12 h).

Cell determined Cell classified
manually as by algorithm as Count Nature of err01

Yeast Double yeast 2 Single cell segments into two cells

Elongated yeast Double yeast 2
Filament Double yeast 2
Double yeast Yeast 1 Double cell fails to segment
Double yeast Elongated yeast 1
Double yeast Filament 4
Double filament Filament 2
Double yeast Double filament 1 Double cell segments at the wrong location
Double filament Double filament 2

O'SHEA AND WALSH: IMAGE ANALYSIS OF DIMORPHIC YEAST 687

 Figure 7. Frequency distribution of cell properties during the fermentation run. (a) volume, (b) length.

tion of these filters is to remove as many of the single useful in avoiding the erroneous placement of double
cells and true mycelia from the original image (Image cells in the single cells image. The effectiveness of Filters
C , Fig. 3c) at the earliest possible stage of the algorithm. 1and 2 is demonstrated by the fact that 84% of the single
The significance of this filtration process is evident from cells in the 12-h sample are isolated in this manner. In
Table VII, which demonstrates that the application of addition, for the same sample, 68%of the cells in Image
the segmentation operations is very time consuming. F (“possible” double cells) are finally classified as true
However, the upper bound of the single cell area and double cells.
the lower bound of the mycelium area must be chosen The greatest difficulties in the development of an
to minimize the number of double cells that would be automatic classification algorithm for the strain studied
incorrectly classified. The use of a perimeterkonvex pe- were encountered with filaments and double filaments.
rimeter filter (Filter 2 of Table 11) was found to be very The lack of a distinct concavity in these cells frequently

688 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 51, NO. 6, SEPTEMBER 20, 1996
resulted in the segmentation of cells into three or more successful characterization of the effects of cell mor-
subunits at points of slight concavity. A second difficulty phology on the filtration behavior of K. marxianus
with the classification of filaments and double filaments broths, a topic which is currently being studied in this
was in distinguishing such cells from similar cells (typi- laboratory. In addition, we intend to use the algorithm
cally double yeast cells) when relying solely on gross to study the effect of cell morphology on the rheological
measures such as feret length and feret width. Slight behavior of K. marxianus broths and also to use continu-
curvature of a filament means that feret width is an ous culture conditions to examine the effect of dilution
inappropriate measure of the mean cell width and, in the rate on the morphological character of K. marxianus.
case of double filaments with an L-shaped morphology,
feret length and width are both inappropriate parame-
References
ters with which to estimate the true cell geometry. In
addition, cells possessing such geometries tended to ov- Cox, P. W., Thomas, C. R. 1992. Classification and measurement of
ersegment. The most suitable way to measure the neces- fungal pellets by automated image analysis. Biotechnol. Bioeng.
39 945-952.
sary geometric parameters proved to be the skeletoniza- Hill, G. A., Robinson, C. W. 1988.Morphological behavior of Sacchar-
tion of the cells in a manner similar to the mycelia. The omyces cerevisiae during continuous fermentation. Biotechnol. Lett.
inability to segment such double filaments necessitated 1 0 815-820.
an alternative means of classification, and the parame- Huls, P. G., Nanninga, N., van Spronson, E. A., Valkenburg, J. A. C.,
ters chosen in Table IV to distinguish between filaments, Vischer, N. 0. E., Woldringh, C. L. 1992. A computer-aided mea-
suring system for the characterization of yeast populations combin-
double filaments, and mycelia were determined by man- ing 2D-image analysis, electronic particle counter, and flow cytome-
ual observation of a large number of cells. try. Biotechnol. Bioeng. 39: 343-350.
The development of an automatic image analysis Kieran, P. M., Malone, D. M., MacLoughlin T.F. 1993. Variation of
method to characterize a cell population containing such aggregate size in plant cell suspension ’ atch and semi-continuous
a wide variety of cell types must involve an element cultures. Trans. IChemE. 71(C1): 40- 5.
Ku, M. A., Hang, Y. D. 1992. Productior ?f yeast lactase from sauer-
of compromise. The system devised had to be truly kraut brine. Biotechnol. Lett. 1 4 925- 28.
automatic so that the routine could be run overnight McNeil, B., Kristiansen, B., Seviour, R. J. 1989. Polysaccharide pro-
on a range of samples from a fermentation. Some reduc- duction and morphology of Aureobmidium pullulans in continuous
tion in accuracy results from the use of a 200X magnifi- culture. Biotechnol. Bioeng. 33: 1210-1212.
cation level. On a sample of 100 single cells (mainly Miller, T. M. 1959. The use of dinitrosalicylic acid reagent for the
determination of reducing sugars. Anal. Chem. 31: 426-428.
yeast and elongated yeast cells), the difference between Packer, H. L., Keshavarz-Moore, E., Lilly, M. D., Thomas, C. R. 1992.
the measured cross-sectional area and the estimated Estimation of cell volume and biomass of Penicillium chrysogenum
cross-sectional area (assuming an ellipse shape) was cal- using image analysis. Biotechnol. Bioeng. 39 384-391.
culated as 6% at 200X magnification and 4% at 400X Peng, C.-S., Chen, T.-L. 1994. Influence of n-hexadecane on penicillin
magnification. In addition, subtle differences in the fermentation with Penicillium chrysogenurn:The morphological ef-
fect. Biotechnol. Tech. 8: 773-778.
boundaries of small cells are lost at the lower magnifica-
Pons, M. N., Vivier, H., Rtmy, J. F., Dodds, J. A. 1993. Morphological
tion, rendering segmentation more difficult. This is off- characterization of yeast by image analysis. Biotechnol. Bioeng.
set, however, by the increase in speed arising from the 4 2 1352-1359.
reduced number of original gray images required to Priede, M. A., Vanags, J. J., Viesturs, U. E., Tucker, K. G., Bujalski,
perform a statistically meaninful analysis. W., Thomas, C. R. 1995.Hydrodynamic, physiological, and morpho-
logical characteristics of Fusarium moniliforme in geometrically
There is an obvious element of overlap in the catego- dissimilar stirred bioreactors. Biotechnol. Bioeng. 48:266-277.
ries chosen to describe the cell population. There is no Reeslev, M., Jensen, B. 1995. Influence of Zn” and Fe3+on polysac-
sharp cutoff point between a “long” elongated yeast charide production and myceliudyeast dimorphism of Aureobmi-
cell and a “short” filament, for instance. However, it is dium pullulans in batch cultivations. Appl. Microbiol. Biotechnol.
felt that the separation of the cells studied into six dis- 4 2 910-915.
Reichl, U., King, R., Gilles, E. D. 1992. Characterization of pellet
tinct categories facilitates a good insight into the nature
morphology during submerged growth of Streptomyces tendae by
of the dynamic growth of a population containing dimor- image analysis. Biotechnol. Bioeng. 39 164-170.
phic cells. The image analysis method described above Rouwenhorst, R. J., Ritmeester, W. S., Scheffers, W. A., van Dijken,
certainly provides a much more comprehensive picture J. P. 1990. Localization of inulinase and invertase in Kluyveromyces
of a fermentation than the simple reporting of popula- species. Appl. Environ. Microbiol. 56: 3329-3336.
San-Blas, G., San-Blas, F. 1984. Molecular aspects of fungal dimor-
tions in terms of the percentage of yeast and filamentous
phism. Crit. Rev. Microbiol. 11: 101-127.
cells (Reeslev and Jensen, 1995; Walker and O’Neill, Schwan, R. F., Rose, A. H. 1994. Polygalacturonase production by
1990). Although the algorithm has only been tested on Kluyveromyces marxianus: Effect of medium composition. J. Appl.
the strain described, it should be capable of categorizing Bacteriol. 76: 62-67.
(with minimal tuning of parameters) any dimorphic Thomas, C. R. 1992. Image analysis: putting filamentous microorgan-
isms in the picture. Trends Biotechnol. 10: 343-348.
strain that displays the same range of morphologies as Tucker, K. G., Chalder, S., Al-Rubeai, M., Thomas, C. R. 1994. Mea-
K. marxianus. The algorithm output allows the calcula- surement of hybridoma cell number, viability, and morphology us-
tion of properties such as the surface mean diameter of ing fully automated image analysis. Enzyme. Microb. Technol.
the cell population. This information is essential to the 1 6 29-35.

O‘SHEA AND WALSH: IMAGE ANALYSIS OF DIMORPHIC YEAST 689

Tucker, K. G., Kelly, T., Delgrazia, P., Thomas, C. R. 1992. Fully- NRRLy2415 during fermentation of lactose. J. Chem. Tech. Bio-
automatic measurement of mycelial morphology by image analysis. technol. 4 9 75-89.
Biotechnol. Prog. 8 353-359. Warren, S.J., Keshavarz-Moore,E., Ayazi Shamlou, P., Lilly, M. D.,
Vienne, P., von Stockar, U. 1985. Metabolic, physiological and kinetic Thomas, C. R., Dixon, K. 1995. Rheologies and morphologies of
aspects of the alcoholic fermentation of whey permeate by Kluyver- three actinomycetes in submerged culture. Biotechnol. Bioeng.
omyces fragilis NRRL 655 and Kluyveromyces lactis NCYC 571. 4 5 80-85.
Enzyme Microb. Technol. 7 287-294. Yamasaki, H., Lee, M.-S., Tanaka, T., Nakanishi, K. 1993. Improve-
Walker, G. M., O’Neill, J. D. 1990. Morphological and metabolic ment of performance for cross-flow membrane filtration of pullulan
changes in the yeast Kluyveromyces marxianus var. marxianus broth. Appl. Microbiol. Biotechnol. 3 9 21-25.

690 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 51, NO. 6, SEPTEMBER 20, 1996