DETECTION OF ISOLATED NEMATODES IN CLUTTER

ENVIRONMENTS USING SHAPE FEATURE HISTOGRAMS

Daniel Ochoa, Sidharta Gautama

Telecomunication and Information Processing Department, Ghent University, St Pieternieuwstraat 41, Ghent, Belgium

Boris Vintimilla

Computer Vision and Robotics Center, Department of Electrical and Computer Science Engineering, ESPOL University,

Km 30.5 via Perimetral,Guayaquil, Ecuador

Keywords: feature extraction, segmentation, recognition.

Abstract: We present an approach for detection of isolated Caenohabditis Elegans nematodes in clutter environments.

The method is based on shape feature histograms which describe the distribution of features of second-order

derivative responses of linear image structures. The shape features are able to distinguish isolated from

overlapping nematodes and clutter, thereby improving the automated image analysis of nematode

populations where accurate assessment of shape is needed. An evaluation is performed on a database of

manually segmented images. Shape continuity features proved to have the highest discriminative power.

This is consistent with the morphological structure of this kind of organism. Our experiments suggest that

similar techniques can be used for identification of other linear shaped biological objects.

1 INTRODUCTION

The increasing amount of digital image data in

biological studies requires efficient and robust image

analysis tools to generate accurate and reproducible

quantitative results. In contrast to medical images

where imaging conditions and sampling methods are

highly controlled, biological images are inherently

difficult to analyse because of sample variation,

noise and clutter. Techniques need to be developed

and constantly adapted to specific tasks, which

requires substantial domain knowledge.

In biotechnology industry, one of the most

common procedures in research labs is the

measurement of microscopic structures to

characterize the interaction of organism population

with chemical substances (e.g. the effect of newly

developed pesticides). This task is typically carried

out by a technician who takes a number of

specimens from a sample to measure their length

and width. The new data is then analysed

statistically to find correlations with certain

chemical compounds. As the number of specimens

and the complexity of the analysis rises, manual

processing of these images becomes less of an

option. Automated image analysis can aid in this

process by identifying and measuring the structures

of interest in the images.

In this paper we work on images containing C.

Elegans nematode populations. This microorganism

has a well-described nervous system, and a complete

genome sequence which makes it attractive for

research and development in biotechnology. In an

earlier paper images containing single nematodes are

examined (Fdez-Valdivia et al., 1992). After

background correction, the image is thresholded and

skeletonized, after which contour curvature patterns

are used to identify the head and tail of the

nematode. By means of an interactive detection

procedure, (Palhares et al., 1997) proposed a method

for nematode recognition based on the stylet contour

morphology. A template matching algorithm

compared it with stylets of nematodes of known

taxonomy. In a first step towards classifying

C.Elegans behavioral phenotypes quantitatively,

(Baek et al., 2002) identified motion patterns by

means of a one-nematode tracking system,

morphological operators and geometrical related

features. All these papers relied on features extracted

486

Ochoa D., Gautama S. and Vintimilla B. (2006).

DETECTION OF ISOLATED NEMATODES IN CLUTTER ENVIRONMENTS USING SHAPE FEATURE HISTOGRAMS.

In Proceedings of the First International Conference on Computer Vision Theory and Applications, pages 486-491

DOI: 10.5220/0001365504860491

 SciTePress

from images with single, isolated nematodes,

segmented using a combination of intensity and

morphological based methods. Nematode

populations were studied in (Van Osta et al., 2002).

In this work scale space principles are applied to line

detection instead of intensity thresholding. The use

of anisotropic diffusion to improve the response of

the line detection algorithm is proposed but no

attempt is made to extract single specimens from the

population. This distorts the reporting of shape

measurements of the detected specimens, like

specimen size and width, if overlapping specimens

are regarded as one.

In contrast to previous efforts aimed at

characterizing individual nematodes, we focus on

detecting isolated nematodes in images of

populations. Given the nature of these images, we

study how to extract reliable shape information for

object identification with a restricted amount of

image data, clutter and structural noise. We consider

identification as a necessary step before any post-

processing task, in particular if a computer vision

based software tool is to be incorporated in daily lab

work where accurate measurements need to be

calculated.

This paper is organized as follows. Section 2

discusses the procedure used to detect the initial

lines that approach to nematodes. Shape

characteristics of isolated nematodes are discussed

and measurements proposed in Section 3. Results

are shown in Section 4, and finally conclusions and

further improvements are presented in Section 5.

2 SEGMENTATION USING

RIDGE DETECTION

In general, nematodes in an image can be thought of

as lines of varying width at each point along their

length, wide in the center and narrow near both ends.

Since nematodes appear as narrow valleys or ridges

in the intensity surface we exploit these properties

for segmentation. There has been a considerable

research into linear object detection in the medical

field particularly for vessel/neurite detection, the

reader is referred to (Kirbas and Quek, 2003) for a

survey of line detection algorithms.

In our work, a scale space method for line

detection is used (Steger, 1998). The intensity

surface f(x,y) is locally approximated by a 2-

dimensional second order Taylor polynomial at

every pixel. Ridges are detected by considering

eigenvectors and eigenvalues λ

± of the Hessian

matrix H:

xxy

yyy

⎡

⎤

⎢

⎥

⎣

⎦

(1)

yyxxyyxx

f+)f(f±f+f=λ± 4

−

(2)

Ridge points show a vanishing second order

derivative in the direction of the eigenvector

corresponding to the maximum eigenvalue λ+.

Calculation of the partial derivatives f

, f

and f

done by convolving (*) the image f(x,y) with

(separable) Gaussian derivative kernels:

(3)

(4)

()

(, ) () ()

xx xx

xy g yg x

σσ

πσ

−

(5)

The ridge detector response R is defined as the

value of the maximum eigenvalue λ+ normalized for

a selected scale

and an estimate of the local

contrast h. To approximate h at every point, the

output of a morphological closing (•) was subtracted

from the image:

)),((),(

seyxfyxfh •

−

(6)

hσλ=R

⋅

(7)

Detection is performed by placing a threshold t

on R to select salient ridge pixels. To form line

segments, pixel chaining which groups connected

pixels belonging to the same line segment, is

required. This is done by taking salient pixels and

adding neighbouring pixels which show evidence of

being part of the same linear structures. The

direction of the eigenvector is used and should lie

within an error margin for pixels to be grouped.

We must point out that the aim of the

segmentation step is to extract linear objects. This of

course implies that many non-nematode objects will

also be segmented. In addition, overlapping

nematodes will be segmented as one or several

objects. When performing a shape analysis of the

segmented objects, the noise structures and

overlapping nematodes will introduce errors if shape

statistics are used to describe the population. It is

therefore important to detect isolated nematodes to

DETECTION OF ISOLATED NEMATODES IN CLUTTER ENVIRONMENTS USING SHAPE FEATURE

HISTOGRAMS

487

be able to produce a reliable reporting on the

properties of the population.

The correct estimation of segmentation

parameters (

, t) still poses a problem and has

motivated different approaches to improve initial

segmentation results (Aylward and Bullitt, 2002;

Amri et al., 2005). In (Steger, 1998), a theoretical

model is developed to determine the optimal

parameters based on the expected line width. This

methodology is used in our work, where we have

derived the optimal parameters for the most common

line profiles found in this kind of image (cfr. Fig 1).

For details, we refer to (Steger, 1998).

Res

onse

0 0.5 1 1.5 2 2.5 3 3.5 4

−0.4

−0.35

−0.3

−0.25

−0.2

−0.15

−0.1

−0.05

Square

Parabolic

Triangular

/σ

-w w

wxh ≤

0))/(1( ≤<−+ xwwxh

wxwxh ≤<− 0))/(1(

wxwxh ≤− ))/(1(

Figure 1: (top) Normalized response R as a function of line

width w and scale

; (bottom) Line profiles equations.

We examined triangular and parabolic line

profiles, common in nematodes images by

approximating their 2D shapes with polynomial and

piecewise functions (cfr. Fig.1, bottom plot). Bar

shaped profile was also included as a reference since

it is used in road, neuron and blood vessel detection.

From derived analytic expressions we found that

despite the profile type, R at center line points can be

expressed as a function of w/

. According to our

estimations the best values for detection task range

from w/

. = 1.8 to 2.5 .The threshold t was set

accordingly to the estimated response R (cfr. Fig.1,

top plot).

3 SHAPE CHARACTERIZATION

Even when pixels with strong ridge responses are

connected into line segments, it is still impossible to

say whether a segment correspond to an isolated

nematode. The lack of salient contour points along

nematode body and overlapping make traditional

approaches such as contour and shape-based

methods difficult to apply. Recognition by means of

appearance/shape models on biological images

(Cootes et al., 1995; Hicks et al., 2002) is a complex

task given the small size of the nematodes in the

image, and the lack of stable landmark points.

Moreover, complex motion patterns prevent the use

of linear systems to create a simple shape model.

Although nonlinear systems have been devised

(Twining and Taylor, 2001) the complete range of

nematode body configurations is still far from being

model.

In this paper, we discuss the use of shape feature

histograms to characterize objects. The idea comes

from the analysis of 2D synthetic line profiles. There

is a relationship between the response R, line

contrast h and line width w. In the case of a line of

constant contrast as scale

gets closer to the

nematode width the response R increases. (cfr. Fig

As can be seen in Fig. 2, isolated nematodes

have a high response R in the middle of the

nematode and lower responses towards the ends. For

isolated nematodes, the response R varies smoothly

in the object compared to the responses of noise

structures and overlapping nematodes. Both

properties can be explained by the fact isolated

nematodes tend to have fairly constant contrast and a

continuous contour while noisy structures have an

unstructured shape and abrupt contrast variations.

Figure 2: Response spatial distribution for: free lying

nematode (left column), paraffin (middle column), and

overlapping nematodes (right column).

The Response distribution in overlapping

nematodes is more complex. Not all the nematodes

VISAPP 2006 - IMAGE ANALYSIS

488

in the group necessarily have the same size, so when

two or more nematodes of different sizes overlap the

number of low response values at junction points is

prone to increase. Also, since those locations

constitute saddle points, second order derivatives

tend to zero and so does the response. This effect is

more considerable when overlapping takes place

near the center of the nematodes body. We can use

this behavior to discriminate overlapping from

isolated nematodes.

Since for each line point response encapsulates

both width and contrast information in one number,

it seems logical to presume that the normalized line

detector response R contains valuable shape

information for recognition. We propose to utilize

this information by examining the histogram of the

response R or related features for each segmented

object. The shape of the histogram is then exploited

to characterize shape specific properties of the

object. The approach is related to shape feature

histograms used for content based image retrieval

(Gagaudakis and Rosin, 2002). Two types of

histograms have been examined: 1) a standard 1D

histogram of the response R, and 2) 2D co-

occurrence matrices.

The standard 1D histogram captures the

frequency distribution of the response R over a

segmented object. Three statistical features have

been chosen to summarize the histogram: mean,

variance and skewness. The variance of the

histogram is useful since variation of the response R

in isolated nematodes is gradual so they are expected

to show a smaller dispersion compared to noisy

structures. The skewness of the histogram is

measured using the third central moment. Skewness

is useful since almost two-thirds of the total length

of the nematodes have approximately the same

width. Therefore it is reasonable to expect that the

histogram for isolated nematodes may exhibit one

peak biased towards the right side and a long left

tail. This type of distribution tends to show negative

values of skewness.

To include spatial information, for each point on

every line segment response values were taken on 2

neighbouring points in every line direction. These

values populated an object co-occurrence matrix C.

Every cell of C is an estimate of the joint probability

P that a pair of points will have values z

and z

When response values are close to each other higher

values will accumulate near the main diagonal of C.

),(

jiij

zzPc =

(8)

EDM = ∑ ∑ (i - j)

i j

(9)

Regarding C as a 2D histogram the element

difference moment EDM, can be use to measure

value dispersion from the main diagonal. In our

experiments EDM of order one was calculated.

Typically used in texture analysis this feature

measures quantitatively how closed are intensity

values in neighbouring pixels. Because objects

corresponding to isolated nematodes are supposed to

display continuous transitions EDM will show

smaller values for objects with smooth contours and

contrast variations (cfr. Fig 3). EDM of higher

orders did not give better results and were therefore

not included in this paper.

Figure 3: Response histogram and Co-occurrence matrix

for free lying nematode (top row), overlapping nematodes

(bottom row).

In addition, length and mean response value of

line segments are also calculated. The length is a

simple and appealing feature to apply since one

could think that the longer the segment the higher

the possibility of corresponding to a nematode. The

mean response has been applied before to detect

salient paths in networks of lines (Geusebroek et al.,

2001). In our dataset, it is difficult to establish a

direct relationship between the mean and nematode

structures but we did not discard the feature.

4 EXPERIMENTAL RESULTS

All these histogram features have been computed on

a set of 20 population images corresponding to

juvenile and adult stage nematodes, captured using

the phase contrast microscopy technique. (cfr. Fig

4).

All images have been segmented manually and

line segments have been classified in isolated and

noise categories. Isolated refers to nematodes that lie

DETECTION OF ISOLATED NEMATODES IN CLUTTER ENVIRONMENTS USING SHAPE FEATURE

HISTOGRAMS

489

freely on the agar substract; noise includes

overlapping nematodes and everything that is not a

nematode such as paraffin, eggs, and dirt.

Once all the features were measured for our

dataset, the range between maximum and minimum

value was binned into 100 intervals after which the

median value was used as threshold for binary

classification. The performance of the classifier is

summarized in receiver-operating-characteristic

(ROC) curves, which represent the trade-off between

sensitivity and specificity of a classifier. For each

histogram feature a ROC curve is drawn,

considering true positives isolated nematodes and

counting overlapping nematodes and noise as

negative.

In Fig. 5, ROC curves for isolated nematodes

detection are displayed. As can be observed

invariably element different moment, EDM exhibits

the best performance. This feature better captures

the smooth variation that characterizes nematode

line segments. A 90% true positive detection rate is

achieved for isolated nematodes in juvenile

nematode populations with only 10% false positives

(cfr. Fig.5, top plot). For this data set we found

recognition consistency even when isolated

nematodes display complex motion patterns. False

negative results occur when a nematode intersects

itself. In this case the presence of junction can cause

a drop in the response value. Skewness along with

mean, length and variance performs poorly.

The experiments were repeated on a second set

of adult nematode population images. Now

nematode's internal organs, which are normally

transparent, were made visible. Our aim was to test

our features in very poor conditions when structural

noise covers most of the nematode body. We found

that intensity variations inside the nematode's body,

particularly in the sections corresponding to the

digestive and reproductive systems negatively

affected the continuity of response values and

consequently the trade off between true and false

detection. (cfr. Fig.5 bottom plot).

The results show that under controlled conditions

it is possible to identify structures of interest by

measuring perceptual clues like smoothness of shape

indirectly without having to fit a specific shape

model to the image.

Figure 4: Juvenile (left), Adult (right) nematodes.

Figure 5: ROC for isolated nematode classification:

normal contrast (top), high contrast (bottom).

5 CONCLUSIONS

In these experiments we have demonstrated that

distributions of second-order derivative responses

are useful to determine shape characteristics of

linear structures applied to nematode detection.

Tests carried out on manually segmented images of

nematode populations show that shape continuity

related features prove to have the most

discriminative power. It seems promising for

VISAPP 2006 - IMAGE ANALYSIS

490

recognition purposes in cases when there is a limited

amount of image data and for biological linear

objects where geometrical configurations are

difficult to model analytically.

Structural noise remains a problem in high

contrast images. Since they can be associated with

the transition between specific parts of the nematode

we are considering extending our methodology to

part detection schemes. The influence of these

features on other types of biological linear structures

such as plant pathogen or cell micro tubular

structures remains as an interesting field for future

work.

ACKNOWLEDGEMENTS

This work was supported by the VLIR-ESPOL

program. Daniel Ochoa is PhD student from VLIR-

ESPOL program. Images were kindly provided by

DevGen corporation, and the Marine Biology

Department of Ghent University.

REFERENCES

Amri Muhammad, Karim Abdul, Roysam Badrinath,

Dowell-Mesfin Natalie M., Jeromin Andreas, Yuksel

Murat, Kalyanaraman Shivkumar, 2005. Automatic

Selection of Parameters for Vessel/Neurite

Segmentation Algorithms. IEEE transactions on

Image processing, 14,1338 – 1350.

Aylward S. R. and Bullitt E., 2002. Initialization, noise,

singularities, and scale in height ridge traversal for

tubular object centerline extraction. IEEE Trans. Med.

Imag., 21, 61–75.

Baek J., Cosman P., Baek J., Feng Z.,Silver J., Schafer W.

2002. Using machine vision to analyze and classify C

Elegant behavioral phenotypes quantitatively. Journal

of Neuroscience Methods, 118, 9-21.

Cootes TF., Taylor C.J., Cooper C.H., 1995. Active shape

model: their training and application. Computer

Vision and Image understanding,6, 38-59.

Fdez-Valdivia J., Perez De la Blanca N., Castilllo P.,

Gomez-Barcina A., 1992. Detecting nematode

features from Digital Images. Journal of Nematology,

24,289-298.

Gagaudakis G., Rosin P. ,2002 Incorporating shape into

histograms for CBIR. Pattern Recognition, 35, 81-91.

Geusebroek J., Smeulders A., Geerts H., 2001. A

minimum cost approach for segmenting networks of

lines. International Journal of Computer Vision, 43,

99-111.

Hicks Y., Marshall D., Martin R.R., Rosin P.L., Bayer

M.M., Mann D.G., 2002. Automatic landmarking for

biological shape model. Proceedings International

Conference on Image Processing, 2. 801-804.

Kirbas Cemil and Quek Francis K.H., 2003. Vessel

Extraction Techniques and Algorithms : A Survey.

Proceedings of the Third IEEE Symposium on

BioInformatics and BioEngineering, 238-246.

Palhares L. ,Bastos R., 1997. A preliminary study for

constructing a computational procedure for nematodes

identification based on morphological aspects.

European Conference for information Technology in

Agriculture, 1,15-18

Steger C., 1998 An unbiased detector of curvilinear

structures. IEEE Trans. Pattern Anal Machine Intell.,

20, 113-125.

Twining C. J., Taylor C. J., 2001.

Kernel Principal

Component Analysis and the Construction of Non-

Linear Active Shape Models. In BMVC 2001, British

Machine Vision Conference.

Van Osta P., Geusebroek J.M., Ver Donck K., Bols L.,

Geysen J., ter Haar Romeny B.M., 2002. The

principles of scale space applied to structure and

colour in light microscopy. Proceedings Royal

Microscopical Society, 37, 161-166.

DETECTION OF ISOLATED NEMATODES IN CLUTTER ENVIRONMENTS USING SHAPE FEATURE

HISTOGRAMS

491