Quantiﬁcation of Stromule Frequencies in Microscope Images

of Plastids Combining Ridge Detection and Geometric Criteria

Birgit M

oller

and Martin Schattat

Pattern Recognition and Bioinformatics, Institute of Computer Science, Faculty of Natural Sciences III,

Martin Luther University Halle-Wittenberg, Von-Seckendorff-Platz 1, 06120 Halle (Saale), Germany

Plant Organelle Shape and Dynamics Lab, Institute of Plant Physiology, Faculty of Natural Sciences I,

Martin Luther University Halle-Wittenberg, Weinbergweg 10, 06120 Halle (Saale), Germany

Keywords:

Plant Cells, Plastids, Stromules, Quantiﬁcation, Segmentation, Ridge Detection, Geometric Criteria, ImageJ.

Abstract:

Plastids are involved in many fundamental biochemical pathways in plants. They can produce tubular mem-

brane out-folds from their surface. These so-called stromules have initially been described over a century ago,

but their functional role is still elusive. To identify cellular processes or genetic elements underlying stromule

formation screens of large populations of mutant plants or plants under different treatments are carried out and

stromule frequencies are extracted. Due to a lack of automatic methods, however, this quantiﬁcation is usually

done manually rendering this step a main bottleneck in stromule research. Here, we present a new approach

for quantiﬁcation of stromule frequencies. Plastids are extracted from microscope images using local wavelet

analysis over multiple scales combined with statistical hypothesis testing to resolve competing detections from

different scales. Subsequently, for each plastid region evidence for the existence of stromules in its vicinity is

investigated applying ridge detection techniques and geometric criteria. Experimental results prove that our

approach is suitable to properly identify stromules. Even in microscopy images with a high noise level and

distracting signals extracted stromule counts are comparable to those of biological experts.

1 INTRODUCTION

Fluorescent proteins have developed into an impor-

tant cell biological tool. They have delivered evidence

that membrane bound compartments in eukaryotic

cells not only possess speciﬁc biochemical proper-

ties, but also speciﬁc shapes and numbers, and often

exhibit striking dynamics. Life cell imaging revealed

that these characteristics can drastically change in re-

sponse to stress. Despite these clear and often strong

responses it is still largely unknown how such chan-

ges support compartment function and how they are

integrated into the cellular regulatory network.

Speciﬁcally interesting membrane bound com-

partments in this regard are plastids. Plastids are uni-

que to plants and are involved in many fundamental

biochemical pathways such as photosynthetic carbon

ﬁxation, which provides us with oxygen as well as

food. Shapes of plastids vary between different plant

tissues as well as plant species, but range usually from

ellipsoid to almost perfectly spherical (Fig. 1A, B). In

response to stimuli such as stress plastids can form

surface membrane out-folds (Fig. 1C), so-called stro-

mules (K

ohler and Hanson, 2000). Although stromu-

les are known to the scientiﬁc community for over a

century and have been intensively studied by the use

of ﬂuorescent proteins for 20 years now, we still can

only speculate about stromule function and regulation

(Hanson and Hines, 2018).

An often used strategy to identify cellular proces-

ses or genetic elements important for stromules are

screens in which larger populations (> 1000) of mu-

tant plants are tested for altered stromule frequencies.

The stromule frequency (or SF%) is deﬁned as the

percentage of plastids carrying at least one stromule

(see also (Schattat and Kl

osgen, 2011)). In addition

to genetic screens, screenings of different chemical

inhibitors, hormones and peptides as well as of bio-

tic or abiotic stresses for their ability to induce chan-

ges to the basic SF% are a common tool to study

stromule regulation (see (Schattat and Kl

osgen, 2011;

Gray et al., 2012) for examples). In order to reliably

assess SF% of a single plant sample, such as a leaf,

three to six randomly chosen spots of the sample are

imaged, capturing up to 1500 plastids (precise num-

ber varies between plant species and tissues) of which

the number of plastids with and without stromule has

to be counted (e.g., (Schattat and Kl

osgen, 2011)).

Möller, B. and Schattat, M.

Quantiﬁcation of Stromule Frequencies in Microscope Images of Plastids Combining Ridge Detection and Geometric Criteria.

DOI: 10.5220/0007390300380048

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 38-48

ISBN: 978-989-758-353-7

Figure 1: Sample image data: A, B) typical individual plastids lacking any stromules, A) four plastids exhibiting ellipsoid

shapes (origin: upper leaf epidermis of Arabidopsis thaliana), B) four plastids being almost perfectly spherical (origin: lower

leaf epidermis of Nicotiana benthamiana), B*) plastid exhibiting a very short spike-like stromule. C) Plastids exhibiting a

straight (C*), curvilinear (C**) or branched (C***) stromule of intermediate length. D, E) representative sample images of

ﬂattened z-stacks of D) lower leaf epidermis (N. benthamiana) and E) upper leaf epidermis (A. thaliana). Insets (D*, D**, E*

and E**) show details of inhomogeneous image background, spurious ﬂuorescence signals, and reﬂections.

To record the shape of plastids, ﬂuorescence pro-

teins are targeted to the plastid stroma (K

ohler et al.,

1997; Marques et al., 2004) and imaged by ﬂuores-

cence microscopy. Although automated microscope

setups can easily generate large and representative

image data sets, there is currently no computer-based

tool available, which could assist with stromule quan-

tiﬁcation. Thus, the extraction of stromule frequen-

cies still mainly relies on visual judgment (Schattat

and Kl

osgen, 2011; Erickson et al., 2017; Kumar

et al., 2018) and makes the assessment the current

bottleneck in many stromule related projects. The-

refore a tool for automated stromule frequency quan-

tiﬁcation would greatly support the advancement of

stromule related research.

The automatic extraction of stromules from ﬂu-

orescent microscopy images faces great challenges

with regard to morphological properties and the over-

all appearance of stromules, their dynamic nature as

well as concerning 3D image quality. Stromules exhi-

bit a notable morphological diversity. In microscopy

images stromules of medium and large size mostly

Quantiﬁcation of Stromule Frequencies in Microscope Images of Plastids Combining Ridge Detection and Geometric Criteria

appear as long thin structures (Fig. 1C), while short

stromules often subsume only a few pixels in length

and appear as small spikes (Fig. 1B

∗

). In addition,

long stromules can be bent, kinked or even branched

(Fig. 1C). Speciﬁcally under conditions inducing stro-

mule formation, stromules can be highly dynamic and

can change their shape as well as position in a matter

of seconds. This can cause problems during z-stack

recording, when stromules have time to move bet-

ween consecutive frames, leading to imaging artefacts

such as duplication of the respective stromule. Howe-

ver, recording of z-stacks spanning the entire cell is

crucial for estimating SF% due to the central vacuole

in plant cells and the rather dispersed nature of plas-

tids and stromules. At the same time it is essential to

image a larger number of cells in a sufﬁcient quality

and resolution, allowing not only to cover many plas-

tids, but simultaneously allowing the detection of thin

stromules reliably. Hence, the analysis of plastids and

stromules demands for relatively fast 3D imaging data

of high resolution.

Standard automated wide ﬁeld ﬂuorescence

systems meet these necessities and are in some

regards superior to confocal laser scanning setups,

particularly concerning speed at higher image resolu-

tions. However, wide ﬁeld ﬂuorescence images may

contain distracting intensity signals from out of focus

planes. This prevents the application of standard

3D image analysis strategies to such z-stacks and

necessitates the projection into 2D images. For

creating those projections no standard maximum

projection method seems feasible, instead extending

depth of ﬁeld software such as CombineZP

(as

described in (Schattat and Kl

osgen, 2009)) has to

be used to create ﬂattened 2D images (Fig. 1D, E).

Despite harbouring the biological features of interest

these extended wide ﬁeld ﬂuorescence images often

suffer from imaging and projection artefacts such

as inhomogeneous background and reﬂections on

cell walls (Fig. 1D

∗

, D

∗∗

). Moreover, the artefacts

often share signiﬁcant similarities with stromules

(Fig. 1D

∗

, D

∗∗

, E

∗

, E

∗∗

) and although they render

the task of developing robust and automated image

analysis tools challenging, the beneﬁt of faster

imaging speed at higher resolution makes wide ﬁeld

ﬂuorescence images still the preferred choice for

high-throughput work.

In this paper, we present a new image analysis

workﬂow for quantifying stromule frequencies from

wide ﬁeld microscopy images facing these challen-

ges. Plastids are robustly detected adopting the ap-

The CombineZP website is dead since August 2017,

see https://en.wikipedia.org/wiki/CombineZ for details.

proach of (Greß et al., 2010) based on local wavelet

analysis in a multi-scale framework. For localizing

stromules in the vicinity of segmented plastids we ap-

ply ridge detection techniques to extract curvilinear

segments as candidates for stromules. Subsequently,

biological characteristics of stromules are exploited

to deﬁne a set of geometric criteria suitable to distin-

guish between curvilinear segments referring to true

stromules and segments originating from distracting

image structures or reﬂections. Experimental results

on images of different types of plastids prove that our

new workﬂow is well-suited to ease the assessment

of stromule frequencies even from challenging image

data and marks an important progress towards fully

automatic high-throughput analysis in this ﬁeld.

The approach has been implemented in Java

as part of MiToBo, a toolbox for processing and

analyzing microscopy images (M

oller et al., 2016),

and the software is publically available. MiToBo

seemlessly integrates into the widely used image

analysis software ImageJ/Fiji (Schindelin et al., 2012;

Rueden et al., 2017) and features its own ImageJ

update site which grants direct access to all functio-

nality and in particular to the new stromule analysis

pipeline. MiToBo is open-source and released under

GPL version 3.0, the source code of MiToBo and

the new analysis pipeline, respectively, are available

from MiToBo’s website

and from Github.

The remainder of the paper is organized as fol-

lows. In Section 2 we review common techniques for

the segmentation of cells and nuclei with similar cha-

racteristics like the plastids in our application, and for

the extraction of curvilinear structures showing paral-

lels to stromules. Our approach is outlined in Section

3 focusing on the new stromule detection method, be-

fore results and a conclusion are given in Section 4

and Section 5, respectively.

2 RELATED WORK

The task of segmenting plastid regions from 2D mi-

croscopy images is very speciﬁc, hence, establis-

hed techniques for tackling this problem do not ex-

ist. Yet, the visual appearance of plastids in ﬂuores-

cence microscopy images and their shape characteris-

tics (Fig. 1) are comparable to those of small cells and

nuclei. Accordingly, the problem of extracting plas-

tid regions from 2D ﬂuorescence microscope images

is deeply linked to the segmentation of such objects

(Chen et al., 2013; Buggenthin et al., 2013; Xing and

Yang, 2016) as well as to the detection of particles and

MiToBo website, http://mitobo.informatik.uni-halle.de

BIOIMAGING 2019 - 6th International Conference on Bioimaging

larger spot-like structures on the cellular level (Bas-

set et al., 2015). The variety of techniques applied

for these problems is manifold, ranging from global

or local image binarization in combination with mor-

phological post-processing and often also a watershed

transformation for separation of touching objects, to

elaborate segmentation methods like graph cuts (Qi,

2014), levelsets (Bergeest and Rohr, 2012), and re-

cently also techniques of deep learning (Kraus et al.,

2016). The detection of spot-like structures (Smal

et al., 2010) is often tackled with binarization and sub-

sequent morphological post-processing, speciﬁc mor-

phological operators like top hat, or h-dome transfor-

mations. Also frequency-based methods like wavelet

analysis are common (Olivo-Marin, 2002).

Stromules are thin protrusions of varying length

emenating from the surface of plastids. In the ima-

ges of our application domain they appear as small

curvilinear segments. The only published attempts to

detect stromules from microscopy images have focu-

sed on detecting and tracking individual stromules in

confocal microscope time series of individual plastids

with the aim to identify stromule to microtubule inte-

raction events (Kumar et al., 2018). Quantiﬁcations of

stromule frequency in the same study were performed

manually highlighting the absence of an available tool

to quantify stromule frequency. In general stromules

show a signiﬁcant structural similarity to objects like

vessels in retinal images (Fraz et al., 2012), streets in

aerial imagery (Salahat et al., 2015), or roots in mini-

rhizotron images (Zeng et al., 2008). For the segmen-

tation of such structures various model-based approa-

ches have been devised. Most of the time these aim at

enhancing vessel-like structures based on model as-

sumptions about their intensity proﬁles and structu-

ral appearance in images (Frangi et al., 1998; Staal

et al., 2004; Moghimirad et al., 2012). The met-

hods are often combined with subsequent segmenta-

tion techniques like thresholding. Some approaches

(Steger, 1998) directly extract curvilinear structures

by integrating initial localization of structures on the

pixel level and subsequent linking procedures consi-

dering local neighborhoods.

3 METHODS

The aim of this work is to efﬁciently extract stromule

frequencies SF% from microscopy images. To this

end we seek to count the total number of plastids in an

image from which at least one stromule emanates and

which yields the basis for determining SF%. Hence,

there is no need for accurately segmenting stromules

in their full length. Rather, we focus on ﬁnding lo-

cal evidence of stromules in terms of thin curvilinear

structures with coherent geometric properties in the

vicinity of detected plastid regions.

Our workﬂow (Fig. 2) is separated in two phases

and decouples plastid detection and identiﬁcation of

stromule parts. In the ﬁrst phase (yellow box in Fig. 2)

plastid regions are extracted from given input images

adopting the method for robust detection of spot-like

structures using wavelets from (Greß et al., 2010). In

the second phase (green box in Fig. 2) each region is

further examined to determine if there is evidence for

the corresponding plastid to form a stromule or not.

3.1 Plastid Detection

Plastids appear in the ﬂattened 2D images as circu-

lar to elliptical objects which on average show brig-

hter intensities than the local background. For seg-

menting them from a given input image we adopt the

approach of (Greß et al., 2010) aiming to extract spot-

like and salient local regions from ﬂuorescence mi-

croscopy images. The key idea of the method is to

extract candidate regions by calculating and threshol-

ding wavelet coefﬁcients over multiple scales. Com-

peting hypotheses for the same image location from

different scales are resolved by statistical testing. Be-

low we only brieﬂy outline the basics of the method,

for further details refer to (Greß et al., 2010).

Initially, the gray-scale input image is iteratively

smoothed with steadily increasing mask sizes re-

sulting in images of successively decreasing resolu-

tion. Wavelet coefﬁcients are extracted as differences

between pairs of smoothed images of adjacent sca-

les. The coefﬁcient images are thresholded applying

a manually selected threshold, and foreground com-

ponents are extracted from the binary images to iden-

tify locally striking intensity patterns as candidate re-

gions for plastids. Since the thresholding is applied

independently to each coefﬁcient image, for a single

image location competing plastid region hypotheses

from different scales may result.

To select a single region as ﬁnal segmentation re-

sult from each set of competing candidate regions

statistical testing is used. It evaluates if a region is

more likely related to a real plastid or originating from

image noise. Given the null hypothesis that the local

region originates from noise, this provides us with a

statistical signiﬁcance, i.e. a p-value, for each candi-

date region. The ﬁnal detection result for each set

of competing regions is then given by the candidate

region associated with the smallest p-value, i.e. ha-

ving the smallest probability to originate from image

noise. Together with all regions distinctly detected at

an image location this yields the ﬁnal set of plastid

Quantiﬁcation of Stromule Frequencies in Microscope Images of Plastids Combining Ridge Detection and Geometric Criteria

Input Image

Plastids

Phase I: Plastid Detection Phase II: Stromule Identification

Region Expansion

Ridge Detection

Candidates

Stromule

Geometric

Criteria

Multi−Scale

Wavelets

Binarization &

Statistical Testing

Plastid &

Stromule

Counts,

SF%

Figure 2: Overview of our approach: in the ﬁrst phase (yellow box) plastid regions are extracted from input images, in the

second phase (green box) curvilinear segments are detected in the vicinity of each plastid region and various geometric criteria

are applied to classify them as referring to a stromule or originating from image artefacts and distracting signals.

regions forming the basis for the second step of stro-

mule identiﬁcation.

3.2 Stromuli Detection

If a stromule emenates from a plastid a thin and cur-

vilinear structure is supposed to appear in the vicinity

of the plastid region. We use a ridge detection method

to localize candidates for such structures (Sec. 3.2.2).

Due to the noise level in the ﬂattened 2D images,

inhomogeneous image background, spurious signals,

and reﬂections, the ridge detection often yields a sig-

niﬁcant number of false-positive detections. Hence,

we apply a collection of task-speciﬁc geometric cri-

teria (Sec. 3.2.3) to exclude erroneously detected seg-

ments and decide about the presence of at least one

stromule for each plastid region extracted in the ﬁrst

phase of our pipeline (Sec. 3.1). Since sometimes

outstandingly bright stromule parts are falsely inclu-

ded in a plastid region during plastid detection, a pre-

ﬁltering step (Sec. 3.2.1) identiﬁes such regions and

applies complementary heuristics for stromule vali-

dation to these speciﬁc regions.

3.2.1 Pre-ﬁltering of Plastid Regions

The plastid detection usually extracts the outlines of

plastids accurately and yields compact and circular to

elliptical regions. Very bright stromules, however, are

sometimes detected as integral part of a plastid re-

gion. Then the basic assumptions about appearance

and characteristics of stromules as outlined in the in-

troduction and on which our detection methods relies

are no longer valid. To handle such situations we per-

form a pre-ﬁltering to identify detected regions not

adhering to our model. Such regions are separately

checked for the presence of stromules.

We ﬁrst calculate the solidity of a plastid region as

the ratio of its area and the area of its convex hull. If

this ratio lies below 0.85 both areas signiﬁcantly dif-

fer and further investigations are required. If parts of

a stromule have accidentally been included in a plas-

tid region the shape of the region is locally thin and

elongated. This property can be used as criterion for

discovering such constellations. We extract the ske-

leton of a region under investigation, identify all end

points and extract the corresponding branches from

the end points up to the next branch or end point. All

branches shorter than 5 pixels or with a distance of

less than half of the total branch length between their

end points are discarded. They are too short or form

a circular structure very unlikely for stromules. Each

of the remaining branches is searched for runs of con-

secutive pixels that have a distance smaller or equal

to 4 pixels to the next background pixel in the binary

region image. If the longest of these runs exceeds a

length of 5 pixels the region is partially thin which is

often a clear indication for the presence of a stromule.

3.2.2 Ridge Detection

For detecting curvilinear segments as candidates for

stromule parts we use the ridge detection approach of

Steger (Steger, 1998) with publically available source

code

. Steger proposes models for the proﬁles of 1D

and 2D curvilinear structures in images. Based on

these models he derives criteria for the directional de-

rivatives of the image function which need to be ful-

ﬁlled by pixels along such structures and which serve

as basis for segment extraction. Initially, respective

pixels are found and for each pixel the local orienta-

tion of a potential segment is estimated from the ei-

genvalues and -vectors of the local Hessian matrix.

Individual pixels are then linked to curvilinear seg-

ments considering the consistency of local orientation

and user-deﬁned contrast assumptions. The ﬁnal re-

sult of the ridge detection is given by all curvilinear

segments exceeding a minimal length threshold.

Ridge (Line) Detection Plugin in Fiji by Wagner/Hiner,

https://imagej.net/Ridge Detection, accessed: 12/12/18

BIOIMAGING 2019 - 6th International Conference on Bioimaging

plastid

stromule

vertices

ellipse

A) exit angle

B) overall

distance

D) number of intersections

C) point distance

along ellipse

candidates

estimated

direction

Figure 3: Geometric criteria to ﬁlter stromule candidates:

they need to emanate from the surface almost perpendicular

(A) and be located close to the plastid (B). In case of ellip-

tical plastid regions they tend to start off close to the ellipse

vertices (C) and usually do not cut through a region (D).

3.2.3 Geometric Criteria for Stromule Detection

The ridge detection usually yields a signiﬁcant por-

tion of false-positive stromule candidates, depending

on noise level, image quality and potential artefacts

due to the image acquisition process and the proce-

dure for generating a ﬂattened 2D image. Thus, to

decide if a plastid region forms at least one stromule

we need to further analyze the candidate segments.

According to the ﬁndings about the appearance of

plastids and stromules in our application and as outli-

ned in the introduction, we model each plastid region

as an ellipse to enable the identiﬁcation of stromules

in their vicinity based on geometric criteria (Fig. 3).

Besides general assumptions about the length of a

stromule or its overall distance d

to the plastid re-

gion (Fig. 3B), we mainly assume stromules to com-

ply with two basic geometric criteria:

i) the angle α enclosed by the local normal ~n of

the plastid surface and the direction of a po-

tential stromule must not exceed a threshold θ

∆

(Fig. 3A), and

ii) for elliptical plastids the distance between the

point of origin of a stromule and at least one ver-

tex of the ellipse must not exceed a maximum dis-

tance θ

(Fig. 3C) .

In addition, stromules usually do not cut through a

plastid region (Fig. 3D) which is a property well-

suited to identify false detections due to reﬂections at

the cell walls (cf. Fig. 1D

∗∗

The different criteria outlined above are imple-

mented in our workﬂow as follows. First, we select

from the results of the ridge detection for each re-

gion the subset of segments which are located close

enough to the region to be eligible as a potential stro-

mule part. To this end, we deﬁne a focus area around

each binary plastid region by iteratively expanding the

region via morphological dilation to a maximal ex-

pansion of 20 pixels. In doing so region identity is

preserved, i.e., dilation locally stops as soon as ot-

her regions are touched. Subsequently, all curvilinear

segments are found which overlap with the expanded

region. Segments can be assigned to multiple regions

since a unique decision is not yet possible.

An ellipse is ﬁt to each region to enable the appli-

cation of the geometric criteria. Each point (x, y) on

an ellipse in normal position, i.e., located in the center

of the coordinate system with major and minor axes

of lengths a and b, respectively, and being oriented

parallel to the x- and y-axes of the coordinate system

fulﬁlls the following condition:

ε(x,y) :=









= 0 (1)

For points located inside the ellipse ε(x,y) < 0 and for

points located outside ε(x,y) > 0.

Given these deﬁnitions for each plastid region and

each assigned segment we determine the number of

intersections and classify the pixels of the segment

according to their distance and relative location to the

ellipse. Pixels of a segment located inside the ellipse

or having a distance larger than 4 pixels to the ellipse

are ignored in the following. Since the ellipses of

the plastid regions are usually not in normal position

pixels and ellipses are shifted and rotated to normal

position prior to classifying the pixels. In addition, to

account for discretization effects outer points are re-

quired to fulﬁll ε(x, y) > 1 instead of ε(x, y) > 0. If

there are no such pixels found on a segment it is com-

pletely discarded. Also, if two or more intersections

with the region ellipse are found the curvilinear seg-

ment is no longer considered since stromules usually

do not cross through a plastid region.

For the remaining segments the checks for a small

exit angle of the stromule and a short distance to the

ellipse vertices are applied. To calculate the exit angle

we determine the intersection point of ellipse and seg-

ment. In some cases no intersection point exists, e.g.,

due to discretization or if parts of a stromule could

not be localized by the ridge detection due to a lack

in contrast. Then we ﬁnd the closest points on the el-

lipse and on the segment, respectively. Subsequently,

we compute the local normal direction of the ellipse

at the corresponding point and estimate the orienta-

tion of the segment by computing the orientation of a

line through the neighboring pixels of the closest seg-

ment point (Fig. 3A). Since stromules usually ema-

nate more or less perpendicular from the plastid sur-

Quantiﬁcation of Stromule Frequencies in Microscope Images of Plastids Combining Ridge Detection and Geometric Criteria

face the angle α between both lines must not exceed

a threshold θ

∆

. Finally, for elliptical plastids the mi-

nimal distance between the intersection point or the

point with minimal distance to the ellipse, respecti-

vely, and any of the two vertices of the ellipse is cal-

culated. If this distance lies below a threshold θ

and

there is still a segment part of at least three pixels

length starting from the intersection point which lies

outside of the ellipse a stromule is assumed.

In conclusion, each plastid region for which at le-

ast one of the candidate segments survives all the ge-

ometric checks is assumed to form a stromule. The

stromule frequency (SF%) of an image is ﬁnally de-

ﬁned as the ratio of plastids which form at least one

stromule in relation to all plastid regions detected in

the image.

4 RESULTS

The plastid detection has already been publically re-

leased some time ago as integral part of the MTBCel-

lCounter

, a semi-automatic tool for counting cells

and sub-cellular structures implemented as plugin in

ImageJ/Fiji (Franke et al., 2015) and based on Mi-

ToBo (M

oller et al., 2016). The stromule detection

has been realized as extension for this tool, hence, is

also publically available. The source code is released

under GPL license and available from Github, additi-

onal documentation can also be found on the MiToBo

website

We have tested our approach on a set of images

of the leaf epidermis from two different plants, Ara-

bidopsis thaliana (Fig. 4) and Nicotiana benthami-

ana (Fig. 5). The MTBCellCounter was executed in

an up-to-date Fiji installation on a virtual machine

using a single core of an up-to-date desktop CPU and

2GB RAM, and running Windows 8, 64-bit, as opera-

ting system. Given these settings processing a single

image takes on average a few seconds to approxima-

tely half a minute of time, depending on the chosen

parameter settings (see below), the number of plastids

present in the image, and of course the overall image

quality and noise level. In general, the code is not yet

optimized for efﬁciency and some intermediate steps

might be speeded-up. Also, as each individual plastid

region is processed independent of other regions pa-

rallelization of stromule analysis would be possible.

The MTBCellCounter allows to adjust the para-

meter settings for the plastid detection depending on

MTBCellCounter page, http://mitobo.informatik.uni-

halle.de/index.php/Applications/MTBCellCounter

MiToBo website, http://mitobo.informatik.uni-halle.de

the concrete images at hand. In addition, the three ge-

ometric criteria applied in stromule validation which

consider the number of intersections, the angle bet-

ween surface normal and segment, and the distance to

the ellipse vertices, can independently be enabled or

disabled according to the type of plastids present in a

speciﬁc experiment. Also the thresholds for the max-

imal difference in orientation and the vertex distance

can individually be conﬁgured by the user.

We selected the images of our test set for which

we present results here to show a signiﬁcant variabi-

lity in image characteristics allowing to demonstrate

the general capabilities and ﬂexibility of our appro-

ach. As a consequence the optimal settings for the va-

rious parameters were selected individually for each

image. In common image data sets, however, the vari-

ance among individual images is usually much smal-

ler than in the test data set and the images show more

similar properties. Hence, in a more realistic expe-

rimental setting a common set of parameters for all

images of a data set can be found easier and with mo-

derate effort.

In Fig. 4 results for a section of one of the ima-

ges of the upper epidermis of Arabidopsis thaliana

are shown. In the top row the basis data for stromule

identiﬁcation, i.e., detected plastids (Fig. 4B) and ex-

tracted stromule candidate segments (Fig. 4D) are de-

picted. In this example all plastids have satisfactorily

been detected. From the initial result of the ridge de-

tection in Fig. 4C the high false-positive rate is obvi-

ous. Particularly in the image background several cur-

vilinear structures have been found which are hard to

distinguish from real stromules considering the local

support of the ridge detection. After ﬁltering out seg-

ments not located close to a plastid, however, many of

the false-positives are already eliminated.

In the bottom row of Fig. 4 the outcomes of the

different checks for geometric consistency are visua-

lized. Plastid regions for which at least one stromule

is hypothesized are shown in green color with the cor-

responding stromule parts being colored in yellow (ir-

relevant pixels inside the ellipse or too far away) and

magenta (relevant pixels), respectively. Four of the

seven plastids in this image form stromules, the three

around the center of the image and the one in the

bottom left corner. By applying just the exit angle cri-

terion with θ

∆

= 30

◦

(Fig. 4E) three of these plastids

are correctly identiﬁed, while in the top left corner an

additional plastid is falsely assumed to form a stro-

mule. If in addition also an ellipse distance threshold

of θ

= 3.0 is applied (Fig. 4G) this false-positive de-

tection is eliminated. The inset in Fig. 4G visualizes

this criterion in detail for a different plastid. In orange

and marked with white arrows you can see the ellipse

BIOIMAGING 2019 - 6th International Conference on Bioimaging

Figure 4: Results for images of A. thaliana: A) part of a typical input image; B) detected plastids; C) result of ridge detection;

D) remaining stromule candidates; E) detected stromules with an exit angle less than 30

◦

; F) result with multi-intersection

check enabled in addition (here no effect); G) same result as in F) with additional ellipse distance threshold applied; H) ﬁnal

output with plastids (yellow) and plastids with stromules (red). For additional information on color encodings refer to the text.

vertices which in this case are relatively far away from

the exit point of the potential stromule on the right.

The stromule visible in the bottom left corner cannot

be identiﬁed in this example due to a lack in contrast

and some blurring around its exit point which already

causes the ridge detection to fail here (Fig. 4C).

In Fig. 5 a sample detection result for an image

clip from one of the images of Nicotiana benthami-

ana is shown. Compared to Arabidopsis thaliana the

plastids in Nicotiana benthamiana are usually more

circular, hence, the distance criterion assuming stro-

mule exit points being located close to the ellipse

vertices does not work well in this case. In addi-

tion, reﬂections at cell walls occur more frequently in

these images. The plastid detection again detects all

three plastids in this example (Fig. 5B) very accura-

tely. Only the plastid in the bottom left corner forms

a stromule while the one in the center of the image

is located close to a reﬂection. From the results of

the ridge detection it can be seen that for both, the

reﬂection and the true stromule, candidate segments

are detected. Without applying the check for multi-

ple intersections both are classiﬁed as stromule parts

(Fig. 5D), while enabling the criterion yields the cor-

rect result with only one plastid forming a stromule

here (Fig. 5E, F). The criterion for the distance be-

tween ellipse vertices and exit point could also help

to eliminate the false-positive detection in this speci-

ﬁc case. However, generally enabling the criterion for

images of Nicotiana benthamiana bears the risk to ex-

clude many true stromule candidates due to the shape

characteristics of the plastids.

Our data set of test images subsumes in total 6

images of Arabidopsis thaliana and 6 images of Ni-

cotiana benthamiana. For all 12 images an automa-

tic detection of plastids and stromules was carried out

applying our new workﬂow. Subsequently, the results

were manually checked and partially corrected by a

biological expert yielding an estimate for the accuracy

of automatic quantiﬁcation. In Fig. 6 scatter plots of

automatic and manual counts for plastids (at the top)

and stromules (at the bottom) are shown. As can be

seen from the plots our approach tends to slightly lar-

ger counts for plastids and stromules than resulting

from manual annotation, though, the absolute diffe-

rences are usually smaller than 10. For stromules

slightly larger differences can be observed. This is

mainly due to stromules being much more difﬁcult to

detect than plastids. Particularly distracting intensity

signals like reﬂections or others curvilinear structures

also present in the images may easily result in false

detections. Nonetheless extracted stromule frequen-

cies usually coincide well between manual and auto-

matic counting.

Quantiﬁcation of Stromule Frequencies in Microscope Images of Plastids Combining Ridge Detection and Geometric Criteria

Figure 5: Results for images of N. benthamiana: A) part of a typical input image; B) detected plastid regions; C) result of

the ridge detection; D) detected stromules with one reﬂection (D

∗

) being erroneously classiﬁed as stromule; F) result with

activated multi-intersection criterion which eliminates the false-positive detection (E

∗

); F) ﬁnal output with detected plastids

(yellow) and plastids forming stromules (red).

●

150

200

250

300

150 200 250 300

Manual plastid counts

Automatic plastid counts

●

120

40 80 120

Manual stromule counts

Automatic stromule counts

Figure 6: Result counts for the 12 test images (black: A. tha-

liana, red: N. benthamiana), on top plastids and at the

bottom stromules. On the x-axes the counts manually ex-

tracted by a biological expert are shown, on the y-axes the

counts automatically acquired with our new workﬂow.

In Fig. 7 extracted stromule frequencies for each

of the 12 test images are plotted, the results of the au-

tomatic extraction in red and the ones of the manual

extraction in black. The overall stromule frequen-

cies (SF%) vary signiﬁcantly between the different

images where in this case images of Nicotiana bent-

hamiana (bars on the right) show larger frequencies

than the images of Arabidopsis thaliana (bars on the

left) due to different treatments. The automatically

and manually extracted stromule frequencies are most

of the time comparable with an average difference of

≈ 0.062. For three images (IDs 2, 9 and 10) the dif-

ferences exceed 0.1 with a maximum of ≈ 0.15. Vice

versa, for ﬁve images the difference lies below 0.05.

0.0

0.2

0.4

0.6

1 2 3 4 5 6 7 8 9 10 11 12

Image ID

Stromule frequency SF%

Figure 7: Comparison of automatically (red) and manually

(black) extracted stromule frequencies for the 12 images of

A. thaliana (IDs 1 − 6) and N. benthamiana (IDs 7 − 12).

Given the unquestionable difﬁculty of the overall

task of stromule identiﬁcation in wide ﬁeld ﬂuores-

cence microscopy images, and given the observation

that also within the manual counting results of diffe-

rent human experts usually noticeable variation in the

numbers of plastids and particularly of stromules can

be observed, the results appear very pleasing. Cer-

tainly the quality of the detection depends on the over-

all image characteristics, and large variability within

a set of images renders the task of stromule identiﬁ-

cation harder. Nevertheless, as could be shown by the

experimental results in this study the approach can al-

ready satisfactorily cope with a signiﬁcant amount of

variation and will allow to extract reasonable stromule

BIOIMAGING 2019 - 6th International Conference on Bioimaging

frequencies in many settings. In addition, the overall

time required for checking and post-processing the re-

sults of the new automatic workﬂow will in almost all

cases be signiﬁcantly smaller than the time necessary

for fully manual annotation of plastids and stromu-

les. Instead of manually annotating several hundreds

of plastids and stromules by hand, usually only up

to 15% of the plastids and an even smaller fraction

of plastids with stromules per image requires manual

processing. This allows to extract large and repre-

sentative data sets much more efﬁciently than before

yielding a suitable basis for biological investigations.

5 CONCLUSIONS

The new image analysis workﬂow for the extraction

of stromule frequencies from wide ﬁeld microscopy

images is capable of extracting reasonable quantita-

tive data suitable for biological investigations. Its per-

formance is comparable to those of human experts

while greatly reducing the time requirements. The

necessity for manual intervention is signiﬁcantly re-

duced to a small fraction of the time that would be ne-

cessary for fully manual annotation. Thus, although

the overall workﬂow is not yet fully automatic and re-

lies on manual parameter tuning as well as on manual

validation and post-processing of results, our appro-

ach marks a signiﬁcant improvement over the state-

of-the-art in stromule studies.

Future work will aim to further increase the de-

gree of automation and improve overall computatio-

nal efﬁciency and detection robustness, particularly

with regard to stromules. One possible direction will

be the investigation of machine learning techniques

for robust stromule identiﬁcation particularly in ima-

ges with a high noise level and low quality.

ACKNOWLEDGEMENTS

This work has been supported by core funding of the

Martin Luther University Halle-Wittenberg, Saxony-

Anhalt, Germany, to B. M. and M. S.

REFERENCES

Basset, A., Boulanger, J., et al. (2015). Adaptive spot de-

tection with optimal scale selection in ﬂuorescence

microscopy images. IEEE Trans. on Image Proc.,

24(11):4512–4527.

Bergeest, J.-P. and Rohr, K. (2012). Efﬁcient globally opti-

mal segmentation of cells in ﬂuorescence microscopy

images using level sets and convex energy functionals.

Medical Image Analysis, 16(7):1436 – 1444.

Buggenthin, F., Marr, C., et al. (2013). An automatic met-

hod for robust and fast cell detection in bright ﬁeld

images from high-throughput microscopy. BMC Bi-

oinformatics, 14(1):297.

Chen, C., Wang, W., et al. (2013). A ﬂexible and ro-

bust approach for segmenting cell nuclei from 2D mi-

croscopy images using supervised learning and tem-

plate matching. Cytometry Part A, 83A(5):495–507.

Erickson, J. L., Adlung, N., et al. (2017). The Xanthomo-

nas effector XopL uncovers the role of microtubules

in stromule extension and dynamics in nicotiana bent-

hamiana. The Plant Journal, 93(5):856–870.

Frangi, A. F., Niessen, W. J., et al. (1998). Multiscale vessel

enhancement ﬁltering. In Medical Image Computing

and Computer-Assisted Intervention (MICCAI), pages

130–137. Springer Berlin Heidelberg.

Franke, L., Storbeck, B., et al. (2015). The ’MTB Cell

Counter’ a versatile tool for the semi-automated quan-

tiﬁcation of sub-cellular phenotypes in ﬂuorescence

microscopy images. A case study on plastids, nuclei

and peroxisomes. Journal of Endocytobiosis and Cell

Research, 26:31–42.

Fraz, M., Remagnino, P., et al. (2012). Blood vessel seg-

mentation methodologies in retinal images - a sur-

vey. Computer Methods and Programs in Biomedi-

cine, 108(1):407–433.

Gray, J. C., Hansen, M. R., et al. (2012). Plastid stromules

are induced by stress treatments acting through absci-

sic acid. The Plant Journal, 69(3):387–398.

Greß, O., M

oller, B., et al. (2010). Scale-adaptive wavelet-

based particle detection in microscopy images. In

Meinzer, H.-P., Deserno, T. M., Handels, H., and Tolx-

dorff, T., editors, Bildverarbeitung f

ur die Medizin,

Informatik Aktuell, pages 266–270, Berlin. Springer.

ISBN 978-3-642-11967-5.

Hanson, M. R. and Hines, K. M. (2018). Stromules:

probing formation and function. Plant Physiology,

176(1):128–137.

ohler, R. H., Cao, J., et al. (1997). Exchange of protein

molecules through connections between higher plant

plastids. Science, 276(5321):2039–2042.

ohler, R. H. and Hanson, M. R. (2000). Plastid tubules of

higher plants are tissue-speciﬁc and developmentally

regulated. Journal of Cell Science, 113:81–89.

Kraus, O. Z., Ba, J. L., and Frey, B. J. (2016). Classifying

and segmenting microscopy images with deep multi-

ple instance learning. Bioinformatics, 32(12):i52–i59.

Kumar, A. S., Park, E., et al. (2018). Stromule extension al-

ong microtubules coordinated with actin-mediated an-

choring guides perinuclear chloroplast movement du-

ring innate immunity. eLife, 7:e23625.

Marques, J. P., Schattat, M. H., et al. (2004). In vivo trans-

port of folded EGFP by the DeltapH/TAT-dependent

pathway in chloroplasts of Arabidopsis thaliana. Jour-

nal of Experimental Botany, 55(403):1697–1706.

Moghimirad, E., Rezatoﬁghi, S. H., and Soltanian-Zadeh,

H. (2012). Retinal vessel segmentation using a multi-

Quantiﬁcation of Stromule Frequencies in Microscope Images of Plastids Combining Ridge Detection and Geometric Criteria

scale medialness function. Computers in Biology and

Medicine, 42(1):50–60.

oller, B., Glaß, M., et al. (2016). MiToBo - a toolbox

for image processing and analysis. Journal of Open

Research Software, 4(1):e17.

Olivo-Marin, J.-C. (2002). Extraction of spots in biological

images using multiscale products. Pattern Recogni-

tion, 35(9):1989 – 1996.

Qi, J. (2014). Dense nuclei segmentation based on graph

cut and convexity-concavity analysis. Journal of Mi-

croscopy, 253(1):42–53.

Rueden, C. T., Schindelin, J., et al. (2017). ImageJ2: Ima-

geJ for the next generation of scientiﬁc image data.

BMC Bioinformatics, 18(1):529.

Salahat, E., Saleh, H., et al. (2015). Novel MSER-

guided street extraction from satellite images. In

IEEE Intern. Geoscience and Remote Sensing

Symp. (IGARSS), pages 1032–1035.

Schattat, M. H. and Kl

osgen, R. B. (2009). Improvement

of plant cell microscope images by use of “depth of

ﬁeld”-extending software. Journal of Endocytobiosis

and Cell Research, (19):11–19.

Schattat, M. H. and Kl

osgen, R. B. (2011). Induction of

stromule formation by extracellular sucrose and glu-

cose in epidermal leaf tissue of arabidopsis thaliana.

BMC Plant Biology, 11(1):115.

Schindelin, J., Arganda-Carreras, I., et al. (2012). Fiji: an

open-source platform for biological-image analysis.

Nature Methods, 9(7):676.

Smal, I., Loog, M., et al. (2010). Quantitative comparison

of spot detection methods in ﬂuorescence microscopy.

IEEE Trans. on Medical Imaging, 29(2):282–301.

Staal, J., Abr

amoff, M. D., et al. (2004). Ridge-based ves-

sel segmentation in color images of the retina. IEEE

Transactions on Medical Imaging, 23:501–509.

Steger, C. (1998). An unbiased detector of curvilinear

structures. IEEE Trans. Pattern Anal. Mach. Intell.,

20(2):113–125.

Xing, F. and Yang, L. (2016). Robust nucleus/cell detection

and segmentation in digital pathology and microscopy

images: a comprehensive review. IEEE Reviews in

Biomedical Engineering, 9:234–263.

Zeng, G., Birchﬁeld, S. T., and Wells, C. E. (2008). Au-

tomatic discrimination of ﬁne roots in minirhizotron

images. New Phytologist, 177(2):549–557.

BIOIMAGING 2019 - 6th International Conference on Bioimaging