Fuzzy-enhanced, Real-time Capable Detection of Biological Viruses

using a Portable Biosensor

Pascal Libuschewski

1,2

, Dominic Siedhoff

, Constantin Timm

, Andrej Gelenberg

and Frank Weichert

Department of Computer Science 7, TU Dortmund University, Otto-Hahn-Str. 16, Dortmund, Germany

Department of Computer Science 12, TU Dortmund University, Otto-Hahn-Str. 16, Dortmund, Germany

Keywords:

Fuzzy Logic, Real-time, Mobility, Biosensor, Biological Virus Detection, GPGPU.

Abstract:

This work presents a novel portable biosensor for indirect detection of viruses by optical microscopy. The

focus lies on energy-efﬁcient real-time data analysis for automated virus detection. The superiority of our

fuzzy-enhanced time-series analysis over hard thresholding is demonstrated. Real-time capability is achieved

through general-purpose computing on graphics processing units (GPGPU). It is shown that this virus detec-

tion is real-time capable on an off-the-shelf laptop computer, allowing for a wide range of in-ﬁeld use-cases.

1 INTRODUCTION

Rising numbers of global virus epidemics increase

the demand for infection control (Mairhofer et al.,

2009). It is desirable to deliver diagnoses on-site

and as quickly as possible, in order to prevent further

spread of virus-transmitted diseases (Erickson et al.,

2008). This calls for virus detection devices which are

fast and portable. Such devices can be used e.g. at air-

ports, to answer the question whether or not passen-

gers might propagate contagious diseases from high

risk regions. A novel method providing the basis for a

portable and real-time capable virus detection device

is the so-called PAMONO technique (Plasmon As-

sisted Microscopy of Nano-Size Objects) (Weichert

et al., 2010; Zybin and et al., 2010). It enables selec-

tive detection of different types of nano-objects, in-

cluding but not limited to viruses. In order to be de-

tected, the nano-objects must be immobilized on the

surface of the PAMONO biosensor. In the case of

viruses, this is achieved by preparing the sensor sur-

face with antibodies. Using different antibodies en-

ables distinction of multiple strains of viruses. Con-

versely, it it possible to determine which antibodies

are capable of attaching to a certain kind of virus.

These capabilities are combined in an inexpensive de-

vice, consisting of the biosensor and of a laptop com-

puter for data analysis.

The contribution of this work is twofold: Firstly, a

fuzzy approach for virus detection is presented as an

improvement over the existing real-time PAMONO

data analysis pipeline described in (Siedhoff et al.,

2011). Secondly, it is described how this analysis

pipeline was ported to mobile computers, maintain-

ing its real-time capability.

The paper is structured as follows: Section 1.1

brieﬂy sketches the physics behind the PAMONO

biosensor. Section 2 describes the data analysis

pipeline, while focusing on the fuzzy virus detection

algorithm. Section 3 covers aspects of mobility. Fi-

nally, sections 4 and 5 provide results and discussion.

1.1 PAMONO Biosensor

PAMONO is the acronym for Plasmon Assisted

Microscopy of Nano-Size Objects, i.e. objects on

the nanometer scale are detected using optical mi-

croscopy (Zybin and et al., 2010). As Ernst Abbe

discovered, optical microscopy is diffraction-limited

to imaging objects that must exhibit a minimum ex-

tension of half the wavelength of the employed light

(Abbe, 1873). As a consequence, optical microscopy

lacks the resolving power to directly detect most types

of viruses. PAMONO is an indirect detection method

that bridges the gap between the nano- and microme-

ter scale by utilizing the surface plasmon resonance

effect. This effect is stimulated by illuminating a

thin gold layer (the sensor surface) with a super-

luminescent diode (Figure 1, left). The gold layer is

prepared with antibodies. Viruses are pumped over

169

Libuschewski P., Siedhoff D., Timm C., Gelenberg A. and Weichert F..

Fuzzy-enhanced, Real-time Capable Detection of Biological Viruses using a Portable Biosensor.

DOI: 10.5220/0004230201690174

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 169-174

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: PAMONO biosensor (left), the recorded data (center) and the concept of virus detection (right).

the gold layer within a buffer solution. If a virus

attaches to an antibody, the surface plasmon reso-

nance properties change within a micrometer scale

area around that nanometer scale attachment. This re-

sults in an increased reﬂectance of the sensor surface

in that micrometer scale area. More light from the

diode is reﬂected at the attachment site, which can be

detected using optical microscopy.

The light reﬂected from the gold layer is cap-

tured with a highly sensitive 12-Bit CCD camera chip,

recording a lateral resolution of 1024 × 256 pixels at

a temporal resolution of 30 frames per second. An at-

taching virus manifests as a quickly appearing, faint

spot in this time-series of images (Figure 1, center).

On a per pixel level, this results in a step in the time

series of intensities (Figure 1, right), allowing to dis-

tinguish virus candidate pixels from non-virus pixels.

The PAMONO sensor detects variations in the thick-

ness of the material covering the sensor surface, thus

enabling detection of any type of object that can be

immobilized on that surface. Its high analytic sensi-

tivity comes at the cost of susceptibility to detecting

artifacts, caused e.g. by sensor jitter and air bubbles,

which have to be sorted out by classiﬁcation, cf. Sec-

tion 2. Details on the physics behind the sensor can

be found in the literature (Zybin and et al., 2010).

2 DATA ANALYSIS

This section sketches the existing pipeline for PA-

MONO data analysis (Siedhoff et al., 2011) and in-

troduces a new fuzzy virus detection algorithm (cf.

Section 2.1). This algorithm can replace the existing

detection based on hard thresholding. Comparative

results are given in Section 4.

The intended use-cases of the PAMONO biosen-

sor demand for an automated data analysis that can be

carried out in real-time. The large amount of data and

the demand for online visualization and diagnosis ne-

cessitate stream processing. An analysis pipeline re-

alizing these concepts will be summarized now. On a

high-level, this pipeline (cf. Figure 2) consists of four

steps: A preprocessing step denoises the images to re-

move Poisson- and other noise inherent to the sensor

setup. To this end, a wavelet denoising ﬁlter (Mitter-

mayr et al., 1996) is used. The second step identi-

ﬁes virus candidate pixels by examining the per-pixel

time-series of intensities. A pattern matching algo-

rithm compares the observed time-series (Figure 1,

right) to the model patterns, given as variations of an

ideal step function. Translational differences on the

time axis are handled by a sliding window. Step three

aggregates sets of contiguous virus candidate pixels

to polygons, using the Marching Squares algorithm.

Finally, step four serves to distinguish detections of

actual viruses from artifacts (false positives). To this

end, form factors are computed from the polygons

which are then used as features in an automatic clas-

siﬁcation, separating viruses from artifacts.

In order to achieve real-time capability and en-

ergy efﬁciency on a mobile device, the entire analy-

sis pipeline is implemented using a GPGPU approach

(Timm et al., 2011a).

2.1 Fuzzy Detection

Pattern matching of per pixel time-series is a cru-

cial part in the detection pipeline (Figure 2): Pixels

covered by the surface plasmon resonance effect of a

virus adhesion can be identiﬁed by inspecting their

time-series of intensities. Time-series of pixels af-

fected by a virus adhesion exhibit a characteristic step

function pattern, while those of unaffected pixels con-

tain only noise (cf. Figure 1, right). To ﬁnd virus can-

didate pixels, the observed time-series are matched

to an ideal model pattern. The task of the follow-

ing fuzzy detection algorithm is to turn the resulting

matching scores into a robust classiﬁcation that elimi-

nates false positives and integrates spatial information

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

170

Classification

Fuzzy Detection

Pattern Matching

Preprocessing

Wavelet denoising

Per-pixel time-series

analysis yielding virus

candidacy scores

Separating virus

candidate pixels from

nonvirus pixels by

fuzzy rules on scores

Eliminating false

positives by regarding

shapes of contiguous

sets of candidates

Figure 2: GPGPU Processing pipeline for biosensor data analysis in real-time.

by taking neighboring matches into account.

Let D(x,y,t) be the matching score returned by the

pattern matcher at spatial position x, y on the t-th im-

age of the time-series of images provided by the sen-

sor. The task is to decide, on the basis of D, whether

or not pixel x, y,t is affected by a virus adhesion. In-

stead of taking a binary decision (like in hard thresh-

olding), fuzzy detection assigns to each pixel a degree

of membership to one or more fuzzy sets. The range

of the membership functions for each set is [0,1]. Un-

like probabilities, degrees of set membership can sum

to values larger or smaller than 1. For the PAMONO

application, fuzzy sets are used to distinguish four dif-

ferent classes, named “virus”, “background”, “noise”

and “artifact”. Our approach extends (M

elange and

et al., 2011) to the ﬁeld of detection enhancement.

In a fuzziﬁcation step, each matching score

D(x,y,t) ∈ R

≥0

is mapped to the fuzzy set µ

step

with

range [0,1], using the soft thresholds 0 ≤ π

< π

,π

step

(d) =











1 for d ≥ π

d−π

−π

for π

< d < π

0 else

. (1)

This fuzzy set indicates similarity of the observed

time-series to the model step function. As step-

function-like time-series are caused by viruses, equa-

tion (1) describes membership to the fuzzy set

of pixels affected by viruses. By using addi-

tional application-speciﬁc fuzzy rules, further domain

knowledge about the structure of virus adhesions can

be incorporated. Three examples of such rules are

given, after the required deﬁnitions have been made.

The 3D neighborhood set N of width w, height h

and length l, evaluated at spatio-temporal coordinates

x,y,t is deﬁned as

x,y,t

w×h×l

{

D(x + ˆx,y + ˆy,t +

}

with

ˆx ∈

{

⌈(w − 1)/2⌉, ...,⌊(w − 1)/2⌋

}

ˆy ∈

{

⌈(h − 1)/2⌉, ...,⌊(h − 1)/2⌋

}

t ∈

{

⌈(l − 1)/2⌉, ...,⌊(l − 1)/2⌋

}

The k-th largest element r(k, S) in the set S with k ∈

{

1,.. . ,|S|

}

is deﬁned as

r(k,S) := s

where s

≥ s

. . . ≥ s

≥ . . . s

|S|

Using these deﬁnitions and building upon the fuzzy

rule µ

,π

step

(d) from equation (1), a further rule for the

virus class can be stated. It aims at enhancing the

detection of pixels at the fringe of virus adhesions:

,π

fringe

(d) = d > π

AND (2)

,π

step

(r(15, N

x,y,t

7×7×3

)).

Here the fuzzy operator a AND b is deﬁned as the

minimum min(a, b). This rule is motivated by the fact

that the magnitude of the surface plasmon resonance

effect decreases with distance. Thus the pixels at the

fringe of an adhesion can not be accounted for by hard

thresholding, due to their lower signal-to-noise ratio

causing a mediocre match, i.e. π

< d < π

. Equa-

tion (2) relaxes the issue by promoting pixels with

mediocre matching scores, if they reside in a neigh-

borhood with a sufﬁcient number of virus pixels. This

is the case for pixels on the fringe. The assigned de-

gree of membership is determined from rank-ordered

matching scores of neighboring pixels.

The second fuzzy rule aims at synchronizing dif-

ferent temporal instances of the same virus detection:

,π

temporal

(

) =

AND

(3)

,π

step

(r(1, N

x,y,t

1×1×5

)).

Equations (1) to (3), together with ﬁve more rules

(omitted for brevity), characterize the class of virus-

affected pixels. There are further rules for the classes

“background”, “noise” and “artifact”. As an example,

a rule for the background class is given as:

,π

background

(d) = d < π

AND (4)

NOT µ

,π

step

(r(10, N

x,y,t

7×7×3

)).

The fuzzy operator NOT a is deﬁned as 1 − a for

fuzzy arguments a. This rule states that background

pixels have matching scores smaller than π

and that

their degree of membership to the background in-

creases, depending on its neighborhood: The smaller

the membership of the 10th best neighboring match-

ing score to the set from equation (1), the larger the

membership to the background class.

In a ﬁnal defuzziﬁcation step, all set memberships

are converted into a crisp decision classifying the pix-

els. As the ratio between the virus class and the other

Fuzzy-enhanced,Real-timeCapableDetectionofBiologicalVirusesusingaPortableBiosensor

171

classes is unknown, the crisp decision is made by

choosing the label of the class with the highest degree

of set membership.

With the goal of real-time capability, the algo-

rithm has been parallelized on the GPU: Images and

detections are stored in a ring buffer. As the rank

function r(k,S) is used by multiple fuzzy rules, the set

S is partially sorted to the point where all rank queries

can be answered in constant time. Global memory

accesses have been minimized by exploiting the lo-

cal memory on the GPU. For a spatio-temporal neigh-

borhood of width w, height h and length l, evaluated

on images of size w

img

× h

img

, the number of global

memory accesses for processing one image amounts

to w

img

· h

img

· w · h · l, if no local memory was used.

On the GPU, each local workgroup of size w

grp

×h

grp

has a shared local memory which can be used as a

manual cache. By doing so, the number of global

memory accesses per analyzed image is reduced to

(⌈w

img

grp

⌉ · ⌈h

img

grp

⌉) · (w

grp

+ w − 1) · (h

grp

h − 1) · l.

As an example, a typical task arising in practice is

analyzing time-series of 1024 × 256 images, using a

neighborhood size of 7 ×7× 3. If a workgroup size of

16 × 16 is used, caching in local memory reduces the

number of global memory accesses from 38,535,168

to 1,486, 848.

3 TOWARDS A PORTABLE VIRUS

DETECTION DEVICE

Portability of the virus detection device comprises

two major aspects: Firstly, the sensor must be made

mobile, as brieﬂy covered in Section 3.1. Secondly, it

must be veriﬁed that the analysis pipeline is suitable

for the resource-constrained computing environment

imposed by portable computers, which will be the fo-

cus of the experiment description in Section 3.2 and

of the results presented in Section 4.

Figure 3: The portable virus detection unit.

3.1 Portable PAMONO Sensor

By design, the PAMONO sensor is a small device,

consisting of four major components:

• a gold-layer mounted on a prism,

• a superluminescent diode for illumination,

• a pump circulating a liquid with the specimen and

• a CCD camera to acquire the images.

Battery-driven operation of the components is possi-

ble and they all ﬁt into a portable case (Figure 3).

3.2 Portable Real-time Data Analysis

The second challenge in creating a portable virus de-

tection device lies in conducting the computationally

demanding analysis of the sensor data in real-time,

while being subject to the resource constraints im-

posed by portable computers. The two major con-

straints are limited processing speed and energy con-

sumption. The focus of this work lies on verifying

the real-time capability of the analysis pipeline when

moving from desktop to laptop computers. Minimiza-

tion of its energy consumption by means of design

space explorations (Timm et al., 2011b) has been con-

ducted previously.

The questions to be answered by the experiments

in Section 4 are as follows: In how far does virus

detection performance beneﬁt from using fuzzy rules

instead of hard thresholds? Can the virus detection

pipeline achieve real-time performance on a portable

computer? What is the speedup of GPGPU? Is a state-

of-the-art CPU sufﬁcient for real-time capability?

These questions are answered empirically by ex-

ecuting virus detection tasks in a controlled environ-

ment. Details of the experimental setup are as fol-

lows: The frame rates attained by a desktop system

(Intel Core i7 2600, Nvidia GeForce GTX 480 1GB,

respectively Nvidia GeForce GTX 560 Ti 2GB) were

compared to those of a laptop system (Intel Core i7

620M, Nvidia Quadro NVS 3100M 256MB, Battery

Capacity 8.4Ah at 11.1V). For both systems, the anal-

ysis pipeline was run on CPU and GPU. The size of

the images and the number of viruses were gradually

increased by repeating a prerecorded 64 × 64 × 160

pixel tile along the x- and y-axes. The examined im-

age sizes are 1024 pixels in x-direction and 128, 256,

512 and 1024 pixels in y-direction. For each combi-

nation of desktop/laptop and CPU/GPU it was investi-

gated which image sizes cause a violation of the real-

time condition. This determines the processing hard-

ware that is capable of real-time performance. For

the laptop system, energy consumption was recorded

for the runs on CPU and GPU by utilizing the battery

status information in the sysfs of Linux.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

172

Table 1: Positive agreement (PA) for fuzzy and non-fuzzy

detection as attained for different sources of noise and at-

tenuation factors of the virus signal.

Noise Signal Non-fuzzy Fuzzy

Source Attenuation PA PA

Poisson 1.00 0.995 1.000

Poisson 0.75 0.975 0.979

Poisson 0.50 0.761 0.828

Sensor 1.00 0.917 0.970

Sensor 0.75 0.801 0.873

Sensor 0.50 0.518 0.593

4 RESULTS

The obtained results divide into measurements of

fuzzy detection performance (Section 4.1) and verify-

ing the real-time capability of the overall data analysis

pipeline on a mobile device (Section 4.2).

4.1 Fuzzy Detection Results

The fuzzy detection algorithm was compared to a

non-fuzzy implementation (Timm et al., 2011a) in

order to assess its merit. Data-driven synthesis was

used to transform real datasets into new datasets for

which ground-truth detection results are known. Pos-

itive agreement (PA) (Cicchetti and Feinstein, 1990)

between ground-truth and detected viruses was used

to measure performance. This choice is motivated by

the fact that it penalizes false positives and negatives,

both of which are equally undesirable in this context.

PA is used as a single omnibus index here, i.e. the

corresponding negative agreement is ignored because

the notion of true negatives in not deﬁned this con-

text: A true negative would be an area not marked as

a virus, neither in the ground-truth, nor in the detec-

tion results. Such areas do not relate to discrete events

on the sensor surface and thus can not be counted.

Suitable parameters for the fuzzy and non-fuzzy

detector were found by optimizing PA by means of a

genetic algorithm (GA). This was done for two types

of datasets: The facile datasets were constructed by

synthesizing real-data-based virus instances on a syn-

thetic Poisson noise background. The more chal-

lenging datasets use real sensor data as the source of

noise. This sensor data contains no viruses but be-

sides Poisson noise it includes several further degra-

dations incurred during image acquisition, such as re-

sulting from concussions of the sensor, dust and air

bubbles in the buffer solution. Different attenuations

of the ground-truth virus signal were added to these

two types of noise signal: The virus intensities were

multiplied by 1, 0.75 and 0.5, respectively. This sim-

ulates decreasing virus sizes because the diameter of

a virus linearly affects the observed virus signal inten-

sity (Zybin and et al., 2010).

Table 1 shows positive agreement (PA) for non-

fuzzy and fuzzy detection after GA-based parame-

ter optimization. Fuzzy detection increases PA in all

cases. The facile datasets (in the ﬁrst two lines) do not

beneﬁt much from fuzzy detection because the signal-

to-noise ratio is high enough to make non-fuzzy de-

tection succeed. For halved signal intensity, fuzzy de-

tection raises positive agreement from 0.761 to 0.828

for the Poisson-noise dataset and from 0.518 to 0.593

for the sensor-noise dataset. It thus enables more re-

liable detection of smaller viruses. The average im-

provement for the sensor-noise datasets amounts to

7 percentage points. Using fuzzy rules in a spatio-

temporal neighborhood is a suitable approach be-

cause it enables lower thresholds (virus areas are more

fully covered and noise-corrupted boundaries can be

closed) without adversely affecting the detector’s sus-

ceptibility to noise (false positive pixels due to noise

do not form contiguous islands).

4.2 Verifying Real-time Capability

The camera used in the PAMONO biosensor captures

30 frames per second, thus deﬁning real-time capabil-

ity as processing 30 or more frames per second (fps).

Artiﬁcially limiting the temporal resolution to relax

the real-time constraint adversely affects detection ro-

bustness and is thus not an option. Table 2 shows that

both examined state-of-the-art CPUs violate the real-

time constraint for all image sizes – including the im-

age size of 1024 × 256 pixels, which is the size that

occurs in practice. To meet the real-time constraint,

it is necessary to utilize the GPU. The mobile GPU is

a low-cost model with 16 streaming processors. As

shown in the table, it attains 32.6 fps for 1024 × 256

images, thus meeting the real-time constraint. Larger

images can be processed in real-time if GPUs with

more than 16 streaming processors are used, like the

desktop GPUs. For the laptop computer, the average

speed-up between GPU and CPU amounts to ≈ 3.6x.

Furthermore, using the GPU instead of the CPU re-

duces energy consumption of the overall system by a

factor of ≈ 3.7x. While processing one 1024 × 256

image on the GPU uses 1.56J, the CPU consumes

5.84J (1024 × 512: 2.76J (GPU), 10.52J (CPU) and

1024 × 128: 0.93J (GPU), 3.26J (CPU)).

Table 2 furthermore shows that the employed

streaming algorithms scale almost perfectly since im-

ages with half the size are processed twice as fast.

This linearity seemingly does not hold for the frame

rates attained by the desktop GPUs. The reason is that

Fuzzy-enhanced,Real-timeCapableDetectionofBiologicalVirusesusingaPortableBiosensor

173

Table 2: Frame rates attained for different image sizes and different devices.

Frames per second (fps)

Image size in pixels

1024 × 128 1024 × 256 1024 × 512 1024 × 1024

Desktop

CPU: i7-2600 23 20.1 16.1 11.5

GPU: GTX 560 Ti 60 60 60 26.8

GPU: GTX 480 60 60 60 40.3

Laptop

CPU: 620M 16 8.5 4.6 2.4

GPU: 3100M 60 32.6 16 7.8

the frame rates are capped at the screen refresh rate

(60 Hz). It is not a limitation incurred by the utilized

algorithms.

In summary, it was shown that using our fuzzy algo-

rithm is beneﬁcial for detecting viruses in PAMONO

data: It increases positive agreement of the detection

results to synthetic ground-truth data. The algorithm,

along with the remainder of the processing pipeline,

achieves real-time performance on a portable device.

The utilization of GPGPU techniques is mandatory

because state-of-the art CPUs do not provide sufﬁ-

cient processing power to satisfy the real-time con-

straint. Using the GPU furthermore saves energy.

5 DISCUSSION

With the increasing global spread of human viral in-

fections and the emergence of highly virulent norovi-

ral strains, the availability of fast, reliable and inex-

pensive methods for virus detection is urgently nec-

essary for screening at e.g. airports or in crisis ar-

eas. The proposed “Portable PAMONO Unit” ful-

ﬁlls these requirements. The unit consists of a small

case containing the novel PAMONO biosensor and of

an off-the-shelf laptop computer running specialized

signal analysis software. Besides allowing for ubiqui-

tous availability of virus detection, it accelerates diag-

noses because results are produced in real-time while

viruses attach to the sensor surface.

Future research aims at a further miniaturization

of the “Portable PAMONO Unit” and at running real-

time data analysis on tablet computers and smart

phones. Due to their low cost, a deployment of a large

number of cooperating “Portable PAMONO Units” is

conceivable. A network of such units allows for draw-

ing conclusions about the large-scale propagative be-

havior of pathogens in the human environment.

ACKNOWLEDGEMENTS

Part of the work on this paper has been supported

by Deutsche Forschungsgemeinschaft (DFG) within

the Collaborative Research Center SFB 876 “Provid-

ing Information by Resource-Constrained Analysis”,

project B2. URL: http://sfb876.tu-dortmund.de

REFERENCES

Abbe, E. (1873). Beitr

age zur Theorie des Mikroskops

und der mikroskopischen Wahrnehmung. Archiv f

Mikroskopische Anatomie, 9:413–418.

Cicchetti, D. V. and Feinstein, A. R. (1990). High agree-

ment but low kappa: II. Resolving the paradoxes.

Journal of Clinical Epidemiology, 43(6):551–558.

Erickson, D., Mandal, S., Yang, A., and Cordovez, B.

(2008). Optoﬂuidic, Electrical and Mechanical Ap-

proaches to Biomolecular Detection at the Nanoscale.

Microﬂuid. Nanoﬂuid., 4:33–52.

Mairhofer, J., Roppert, K., and Ertl, P. (2009). Microﬂu-

idic systems for pathogen sensing: A review. Sensors,

9(6):4804–4823.

elange, T. and et al. (2011). Fuzzy random impulse noise

removal from color image sequences. IEEE Transac-

tions on Image Processing, 20(4):959–970.

Mittermayr, C., Nikolov, S., Hutter, H., and Grasserbauer,

M. (1996). Wavelet denoising of Gaussian peaks:

A comparative study. Chemometrics and Intelligent

Laboratory Systems, 34(2):187–202.

Siedhoff, D., Weichert, F., Libuschewski, P., and Timm, C.

(2011). Detection and classiﬁcation of nano-objects

in biosensor data. Microscopic Image Analysis with

Applications in Biology (MIAAB 2011).

Timm, C., Libuschewski, P., Siedhoff, D., Weichert, F.,

uller, H., and Marwedel, P. (2011a). Improving

nanoobject detection in optical biosensor data. Proc.

5th International Symposium on Bio- and Medical In-

formation and Cybernetics (BMIC 2011), 2:236–240.

Timm, C., Weichert, F., Marwedel, P., and M

uller, H.

(2011b). Design space exploration towards a realtime

and energy-aware GPGPU-based analysis of biosen-

sor data. Computer Science – Research and Develop-

ment, Springer, pages 1–9.

Weichert, F., Gaspar, M., Timm, C., Zybin, A., Gurevich,

E., Engel, M., M

uller, H., and Marwedel, P. (2010).

Signal analysis and classiﬁcation for surface plasmon

assisted microscopy of nanoobjects. Sensors and Ac-

tuators B: Chemical, Elsevier, 151:281–290.

Zybin, A. and et al. (2010). Real-time detection of single

immobilized nanoparticles by surface plasmon reso-

nance imaging. Plasmonics, 5:31–35.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

174