ABLE: An Automated Bacterial Load Estimator for the

Urinoculture Screening

Paolo Andreini, Simone Bonechi, Monica Bianchini, Andrea Garzelli and Alessandro Mecocci

Department of Information Engineering and Mathematics, University of Siena, Via Roma 56, Siena, Italy

Keywords:

Image Classiﬁcation, Automatic Urinoculture Screening, Urinary Tract Infections.

Abstract:

Urinary Tract Infections (UTIs) are very common in women, babies and the elderly. The most frequent cause

is a bacterium, called Escherichia Coli, which usually lives in the digestive system and in the bowel. Infections

can target the urethra, bladder or kidneys. Traditional analysis methods, based on human experts’ evaluation,

are typically used to diagnose UTIs, an error prone and lengthy process, whereas an early treatment of common

pathologies is fundamental to prevent the infection spreading to kidneys. This paper presents an image based

Automated Bacterial Load Estimator (ABLE) system for the urinoculture screening, that provides quick and

traceable results for UTIs. Infections are accurately detected and the bacterial load is evaluated through image

processing techniques. First, digital color images of the Petri dishes are automatically captured, and cleaned

from noisily elements due to laboratory procedures, then speciﬁc spatial clustering algorithms are applied to

isolate the colonies from the culture ground and, ﬁnally, an accurate evaluation of the infection severity is

performed. A dataset of 499 urine samples has been used during the experiments and the obtained results are

fully discussed. The ABLE system speeds up the analysis, grants repeatable results, contributes to the process

standardization, and guarantees a signiﬁcant cost reduction.

1 INTRODUCTION

Urinary Tract Infections (UTIs), together with those

of the respiratory tract, are of great clinical relevance

for the high frequency with which they are found in

common medical practice and because of the com-

plications arising therefrom. UTIs can target the ure-

thra, bladder or kidneys and they are mainly caused by

Gram–negative microorganisms, with a high preva-

lence of Escherichia Coli (E. Coli, 70%) — which

usually lives in the digestive system and in the bowel

—, even if clinical cases frequently occur where com-

plicated infections are caused by Gram–positive or

multi–resistant germs, on which the common antimi-

crobial agents are inevitably ineffective, leading to

therapeutic failures.

The urinoculture is a screening test in the case

of hospitalized patients and pregnant women. In the

standard protocol, the urine sample is seeded on a

Petri dish that holds a culture substrate, used to arti-

ﬁcially recreate the environment required for the bac-

terial growth, and incubated at 37

◦

C overnight. After

the incubation, each dish must be examined by a hu-

man expert, adding some more time to the medical

report output. This common situation signiﬁcantly

departs from the needs of clinicians to have results

in quick time, to set a targeted therapy, avoiding the

use of broad–spectrum antibiotics and improving the

patient management

. Moreover, traditional analysis

methods suffer from further problems, such as pos-

sible errors arising in the visual determination of the

bacterial load — due to the skills and the expertise of

the individualoperator —, anddifﬁculties in the trace-

ability of samples and results (Ballabio et al., 2010).

Recently, signiﬁcant improvementsin biology and

medicine applications and decision support systems

(Berlin et al., 2006) have been obtained by using hy-

brid methods, based on a combination of advanced

image processing techniques (Deserno, 2011; Belazzi

et al., 2011), artiﬁcial intelligence tools (Agah, 2014;

Heckerling et al., 2007; Bianchini et al., 2013), ma-

chine learning (Bandinelli et al., 2012), expert sys-

tems, fuzzy logic (Torres and Nieto, 2006), genetic al-

gorithms, and Bayesian modelling (Dey et al., 2010).

In particular, the development of automated tools for

results assessment (screening systems) has attracted

increasing research interest during the last decade, be-

cause of their higher repeatability, accuracy, reduced

staff time (that are the main limiting factors of man-

ual screening), and lower costs (Bourbeau and Lede-

Rapid reporting is crucial, especially when pediatric

patients are involved, since, in this case, the infection symp-

toms are not always speciﬁc, while it is urgent to decide if

an antibiotic therapy is necessary or not and when to start it.

Andreini, P., Bonechi, S., Bianchini, M., Garzelli, A. and Mecocci, A.

ABLE: An Automated Bacterial Load Estimator for the Urinoculture Screening.

DOI: 10.5220/0005687005730580

In Proceedings of the 5th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2016), pages 573-580

ISBN: 978-989-758-173-1

573

boer, 2013). Automated urinalysis devices improve

the capacity of the laboratory to screen more samples,

producing results in less time than by manual screen-

ing. Moreover, the redeployment and lower grading of

staff with the increased turnover and speed of urine

screening, gave economic advantages of automated

screening over manual screening (NHS Purchasing

and Supply Agency, 2011).

Even if some interesting research has been carried

out in recent years for the urinoculture screening (An-

dreini et al., 2015), tracing the state–of–the–art in im-

age processing/AI solutions to the automatic analy-

sis of Petri dishes is difﬁcult, since published results

are often related to subtle variations of the core prob-

lem (that ranges from the classiﬁcation of the infec-

tion type to the evaluation of its severity), pertinent

to various domains (from food and beverage safety

to environmental control and speciﬁc clinical analy-

ses (Ogawa et al., 2012; Clarke et al., 2010; Brugger

et al., 2012; Wang, 2011; Chen and Zhang, 2009)),

and based on different datasets.

In this paper, we propose a tool called ABLE (Au-

tomatic Bacterial Load Estimator), that provides a de-

cision support system for biologists. The system auto-

matically gets dish images from a color camera and,

through a suitable preprocessing phase, removes the

non relevant elements, due to laboratory procedures

(such as labels and written text). Then, ABLE im-

plements spatial clustering techniques to isolate the

colonies from the culture ground, even in presence of

ground disuniformities. Thereafter, using the infor-

mation obtained thanks to the background removal,

the system identiﬁes the infected plates, also estab-

lishing their bacterial load. Finally, ABLE is capable

to reveal the presence of multiple infections grown on

the same dish, to alarm the analyst for contamination.

The ABLE system allows a substantial speedup of

the whole procedure, besides avoiding the continu-

ous transition between sterile and external environ-

ments. The ﬁnal outcomes are directly stored along

with the related analysis records (the image, the type

of infection — unique or multiple —, and the colony

count). Data used during experiments have been pro-

vided by DIESSE Ricerche Srl, Siena. Preliminary

experiments show very promising results.

The paper is organized as follows. In Section

2, the procedure adopted to remove the noisily ele-

ments from images is described. Section 3 presents

the background subtraction system and the method

used to decide if a sample is infected or not, to eval-

uate the bacterial load; the procedure for establish-

ing the number of different infections contemporary

grown on the same plate is also illustrated. Finally,

Section 4 collects experimental results, whereas con-

clusions are drawn in Section 5.

2 PREPROCESSING

In general, the plate–handling process requires some

ancillary data that are added on the plate after the

urinoculture seeding procedure (for instance, a writ-

ten text with the manufacturer name, a label for pa-

tients traceability, etc.). The plates, used in this

work, contain written text on the background for the

type of culture ground (Agar chromID CPS) and the

bioMérieux trademark. Moreover, a label is pasted

underneath with the patient name and a bar code (see

Fig. 1). These elements negatively affect the clas-

siﬁcation process and must be removed. They have

a ﬁxed shape and dimension, but their position can

change from image to image.

2.1 Written Text Removal

To remove the text, a ﬁrst problem to be solved is its

precise localization inside the image. We have used a

template matching approach, based on a sample of the

written text manually extracted from an almost empty

plate (see Fig. 1). Being the text ﬁxed, the manual op-

eration needs to be performed only once, at the very

beginning of the preprocessing phase (thereafter it is

stored in the system). Gradient variations (based on

Sobel ﬁltering (Gonzalez and Woods, 2008) proce-

dure) were evaluated to guarantee independence from

light alterations.

Figure 1: Written text removal scheme.

Normalized Cross Correlation (NCC) (Ahuja and

Tuli, 2013) has been used to detect the text position:

R(x,y) =

∑

x,y



f (x, y) −



(t (x, y) − t)

where n is the number of pixels in t(x,y) (the tem-

plate) and in f(x,y) (the image),

f, σ

, t and σ

are,

respectively, the average and the standard deviation of

f and t. While the text appearance is ﬁxed (and can be

stored), its rotation is not, and must be compensated.

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

574

To this end, different rotated versions of the acquired

template are applied to the image and the best match is

selected. To speed up the process, the template search

area is limited to a subpart of the whole image. In

fact, the text printing process grants some tolerance

limits to the positional variability (e.g., the distance

between the text position and the image center cannot

exceed a certain threshold).

2.2 Label Removal

The next preprocessing step aims at removing the area

occupied by the label, attached under the plate. The

image acquisition device uses back lighting, so that

the light passes through the semi–transparent culture

ground and the label, and the latter absorbs the most

of the light. As a result, the label area is always darker

than the surroundings. To segregate the label, we

use an adaptive threshold obtained by applying the

Otsu’s method (Otsu, 1979) to the image luminance

(Fig. 2 (b)). The binary mask gained after thresh-

olding contains the darkest regions in the plate (some

colonies and the label). A morphological opening is

then used to regularize the mask shape, based on a

circular structuring element, with a diameter slightly

smaller than the shortest label side. In this way, bac-

terial colonies and other artifacts, smaller than the la-

bel, are removed (Fig. 2 (c)). Finally, the minimum

perimeter rectangle of the largest connected compo-

nent is computed, recoveringthe label position. Pixels

belonging to this rectangle are disregarded during fur-

ther processing steps. Moreover, the patients’ name

is blurred by applying a severe smoothing on some

ﬁxed positions relative to the detected label, in order

to work without worrying about privacy issues (Fig. 2

(d)).

3 INFECTED PLATE DETECTION

The main requirement for ABLE has been that of cor-

rectly identifying positive samples. In general, the

number of negative samples is greater than that of

positive samples (more than 60% are negative results)

(Broerm et al., 2011). Moreover, negative samples

have small clinical relevance since, generally, they do

not need further examination. So, for a biological

laboratory, a highly accurate classiﬁcation into pos-

itive and negative cases represents a large workload

reduction. Usually, a plate can be considered posi-

tive if the number of microorganisms per milliliter of

urine exceeds 10

. Our dataset samples were seeded

using a bioMérieux Previ Isola automated agar plate

inoculation system. This device starts from a ﬁxed

(a) (b)

Figure 2: In (a), the original image; in (b), the mask ob-

tained with Otsu thresholding; in (c), the result of morpho-

logical opening; in (d), the red rectangle shows the label,

and the blue rectangles show those parts that were blurred.

point and circularly spreads the urine sample over the

whole plate: more serious is the infection, the greater

is the angle between the starting point and the last

colony grown on the plate. If the angle is wider than

180 degrees, then the sample is considered to be posi-

tive (Rice and Baruch, 2009). To identify the spread–

angle, the bacterial colonies must be segregated from

the culture ground.

3.1 Color Space Analysis

Since a chromogenic medium is used as the ground

seed, the pixel color is one of the most important

feature to distinguish the bacterial colonies from the

background. Therefore, preliminarily, the distribu-

tion of background colors has been analyzed in four

different color spaces (i.e., RGB, HSV, CIE–Lab, and

YCrCb). A supervised training procedure has been

adopted, during which a human expert selected about

80 different regions belonging to the background

and to the foreground. The same sample regions,

extracted from a subset which was not employed in

the testing phase, have been used also for training the

Gaussian mixture models (see Sections 3.2 and 3.4).

The chromatic components of the pixels belonging

to such regions are accumulated to represent the

typical background and foreground chromatic values.

The Dunn’s Index (DI) has been used to give a

quantitative rank (based on the Centroid Linkage

distance and the Centroid Diameter dispersion):

ABLE: An Automated Bacterial Load Estimator for the Urinoculture Screening

575

DI(X) =

min

1≤i≤ j≤k

d(C

)

max

1≤s≤k

{∆(C

)}

d(C

) ,

|+|C



∑

~c∈C

d(~c, µ

) +

∑

~c∈C

d(~c, µ

)



∑

~c∈C

∆(C

) ,

∑

~c∈C

d(~c,µ

)

where X represents the color space where DI is

calculated, ~c is the chromatic vector of each pixel, k

is the number of clusters, d(C

) is a dissimilarity

function between two clusters C

and C

, ∆(C

) is the

mean distance of all the points from the mean, and µ

is the mean of cluster i. DI is higher in the CIE–Lab

space, which indicates a better clustering ability com-

pared with the other color spaces. The same conclu-

sion can be achieved by visually analyzing the color

distribution in the four different spaces (Fig. 3).

(a) (b)

Figure 3: Typical foreground elements are represented with

their own colors, whereas the typical background color is

plotted in blue. The scatter plots in (a), (b), (c) and (d)

represent, respectively, the color distribution in the HSV,

CIE–Lab, RGB and YCrCb color spaces.

3.2 Background Subtraction

Although the acquisition device uses a controlled illu-

mination system, the effect of agar dishomogeneities,

and of light disturbances from the external environ-

ment, produces relevant brightness variations in the

background. Therefore, only the (a, b) chromatic

components of the CIE–Lab color space have been

used for detecting colonies, in order to gain a robust

representation with respect to lighting changes, shad-

ows and local variations. Moreover, the presence of

some particular type of infections (Proteus Mirabilis)

signiﬁcantly changes the culture ground appearance

(see Fig. 4). Consequently, the background clearly

shows two different clusters in the CIE–Lab color

space (as evidenced by the two blue areas in Fig. 3).

(a) (b)

Figure 4: In (a), the typical background color and, in (b), the

background appearance changed by the Proteus Mirabilis

infection.

By analyzing the background color samples by

means of the Mardia and Henze–Zirkler normality

tests (Mardia, 1970; Henze and Zirkler, 1990), it has

become evident that a simple Gaussian model is un-

suitable for modeling the two background clusters

(see the contour lines of the background distributions

in Fig. 5). This explains why a Gaussian mixture

model (GMM) has been adopted to describe the cor-

responding multimodal density functions:

p(θ) =

∑

i=1

N(µ

,Σ

)

where θ = (~µ,

Σ) collects the mixtures parameters,

whereas the i–th vector component is characterized

by a normal distribution with weight Φ

, means µ

and covariance matrix Σ

. The number of mixture

components has been empirically chosen by observ-

ing the data contour lines, while the Expectation–

Maximization (EM) algorithm has been used to es-

timate the mixture parameters. In Fig. 5, the contour

lines estimated by the model are compared with those

obtained using the original data. A similar approach

has been also applied to the infection detection prob-

lem, as explained in Section 3.4.

The background subtraction procedure is shown

in Fig. 6; a Mean–Shift segmentation algorithm is

used to compensate for local background dishomo-

geneities. For each segment, the (a,b) modal values

are compared with the Gaussian mixture models and,

if the posterior probability of the background is the

greatest one, the correspondingsegment is considered

to be part of the culture ground. Some speciﬁc type of

infections cannot be classiﬁed by using chromatic in-

formation only, because their color is very similar to

the background. In this case, spatial features (i.e., ob-

tained with edge detection techniques) must be used

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

576

(a) (b)

Figure 5: In (a) and (b), the contour lines of the two back-

ground models; in (c) and (d), the contour lines estimated

by the mixture model.

to obtain a suitable segmentation performance. In Fig.

7, some results are reported.

Figure 6: Background subtraction scheme.

3.3 Infection Severity Estimation

The angle between the inoculation starting point and

the last colony found on the plate gives an indication

of the infection severity. The plate image is divided

into 64 equiangular sectors. For each sector, the fore-

ground concentration is computed (number of fore-

ground pixels divided by the total number of pixels

in the sector). The sector with the maximum con-

centration is considered as that containing the inoc-

ulation starting point. From the starting sector, the

image is analyzed counter–clockwise (opposite to the

seeding direction) sector by sector, until a not empty

sector is found. This last sector is considered as the

colony proliferation end point. The angle between the

starting point and the end point, in the clockwise di-

rection, is used as a measure of the infection spread–

angle over the plate. If the angle is wider than 180 de-

grees, the sample is considered as positive (infected).

(a) (b)

(e) (f)

Figure 7: In (a), (c), and (e), some original images; in (b),

(d), and (f), the background subtraction results; the two

background classes are respectively colored in green and

dark green.

Unfortunately, our experiments clearly show that the

spread–angle is not accurate enough to predict the in-

fection severity. In fact, even if a colony spreads over

the whole plate, some sectors within the spread–angle

can be actually empty. To compensate this error, we

simply subtract the angular contribution of the empty

sectors from the estimate of the spread–angle. The

algorithm is sketched in Fig. 8.

Moreover, further analyses have been carried out

on the positive samples only, with the aim of distin-

guishing among severely infected plates (≥ 10

) and

moderately infected plates (≥ 10

). To this aim, we

estimate, one sector after the other, the ratio between

the infected area and the sector area.

3.4 Detection of Multiple Infections

Biological laboratories daily examine a huge number

of Petri plates. When a sample contains more than

two infection strains, the plate needs further, speciﬁc,

ABLE: An Automated Bacterial Load Estimator for the Urinoculture Screening

577

(a) (b)

Figure 8: The original image (a) is divided in 64 sectors (b);

in (c), the spread–angle is estimated and, in (d), only the not

empty sectors (in red) are considered.

analysis. Therefore, it is important to quickly detect

the presence of multiple infected plates where every

type of infection is signiﬁcantly present. Our pro-

posed algorithm aims at detecting the following in-

fection classes: E. Coli, Enterococcus Faecalis, KES

group, Proteus, Pseudomonas Aeruginosa, and Mor-

ganella. The last three classes are not well repre-

sented within the dataset and, therefore, since they

all produce yellow colonies, we decided to group

them together in a "yellow infection" pseudo–class.

When more than two classes are contemporary de-

tected on the same plate, the sample has to be con-

sidered as "contaminated". To this end, the CIE–Lab

(a,b) values of the four classes (E. Coli, Enterococ-

cus Faecalis, KES group, yellow class) have been

extracted, using some foreground samples collected

during the initial color space analysis phase. As for

the background subtraction module, only the chro-

matic components have been used. Again, the Mardia

and Henze–Zirkler normality tests indicate that sim-

ple Gaussian models are unsuitable. The contour lines

of each distribution are shown in Fig. 9.

As for the background subtraction module, a

Gaussian mixture model (GMM) has been used to de-

scribe the various probability density functions. After

the background subtraction, a Mean–Shift segmenta-

tion algorithm was applied in order to compensate for

local dishomogeneities. For each segment, the (a,b)

modal values are compared with the Gaussian mix-

ture models of all the classes, and the maximum pos-

terior probability gives the classiﬁer output. It has

(a) (b)

(e) (f)

(g) (h)

Figure 9: In (a), (b), (c) and (d), the contour lines, and in (e),

(f), (g) and (h), the lines estimated with the mixture model

respectively for E. Coli, E. Faecalis, the "yellow class", and

KES.

been noted that, sometimes, there are halo–regions

surrounding colonies. These halo–regions actually

belong to the background but, in some cases, the pres-

ence of groups of nearby colonies changes the back-

groundappearance, producingsome errors in the clas-

siﬁcation output. Moreover, when different types of

infections, with different colors, overlap on the same

plate, their color changes in the melding region, and

this also leads to incorrect classiﬁcations. Since the

unpredictable colors produced in these regions do not

belong to any previously deﬁned model, the posterior

probability in those areas is likely to be low. To im-

prove the performance in such situations, low values

of the posterior probability are used to identify the

uncertainty areas on the plate (see Fig. 10). The ac-

tual number of infections is determined with respect

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

578

to only those regions with a well–deﬁned color.

(a) (b)

Figure 10: In (a), the original image; in (b), in gray, the

uncertainty region, in green, the background, in red, blue,

dark blue and yellow the various infections.

4 RESULTS

The preprocessing algorithms have been tested on

the whole data set comprised of 499 images. Over-

all, written texts have been correctly identiﬁed in the

86,78% of the images (433/499). However, it must

be noted that the actual performance depends on the

level of clutter in each plate. In the case of infected

plates, a relevant part of the visual area is covered by

the infection and the text is only partially visible (see

Fig. 11). As a result, in infected images, the correct

text detection rate is 75,45% (160/212), whereas in

non infected images it is 95,12% (273/287).

Figure 11: Two examples of highly infected plates, where

the written text is signiﬁcantly occluded.

It is worth noting, from a practical point of view,

the text detection is not important for highly infected

plates, because the infection severity mainly depends

on how much the colony is spread on the plate, and

the eventual few undetected text pixels do not inﬂu-

ence the automatic classiﬁcation results. The label re-

moval algorithm also shows very good results: 100%

of the whole dataset of 499 images has been correctly

processed.

After the preprocessing phase, the system performs

the infection severity estimation. The performance

obtained by applying the procedure described in Sec-

tion 3.3 to our dataset (accuracy and confusion ma-

trix

) are reported in Table 1. As we can see, 22 im-

ages were incorrectly classiﬁed. However, it is impor-

tant to note that these images are all false positive, so

that we obtain a sensitivity of 100%. In fact, in our

dataset, a positive sample is never confused with a

negative one (false negative). This is a very desirable

result, since a false positive "only" requires a further

analysis by the human expert, whereas a false nega-

tive could lead to ignore the infection and to expose

the patient to possible risks.

Table 1: Accuracy and confusion matrix obtained by ABLE.

Number of images 499

Correctly classiﬁed 477

Incorrectly classiﬁed 22

Accuracy 95,4%

Confusion Matrix

Positive Negative

212 22

0 265

For the 212 positive samples in our dataset — to be

distinguished between severely infected (≥ 10

) and

moderately infected (≥ 10

) — the obtained results

are shown in Table 2. Although the classiﬁcation ac-

curacy was not astonishing, we can assert that, in al-

most all the cases in which ABLE fails, also the judg-

ments of the human experts (three biologists, in this

case) are mostly discordant.

Table 2: Classiﬁcation accuracy for infected plates (≥ 10

vs. ≥ 10

), and confusion matrix, obtained by ABLE.

Number of images 212

Correctly classiﬁed 176

Incorrectly classiﬁed 36

Accuracy 83,01%

Confusion Matrix

≥ 10

57 8

28 119

Even if the experimental results are actually very

promising, they are devised on a small set of data,

whereas the ABLE system must be experienced in

A confusion matrix allows the visualization of the per-

formance of a supervised learning classiﬁer. Each column

represents the instances in the predicted class, while each

row represents the instances in the actual class. Its name

stems from the fact that it clearly shows if the system is

confusing two classes (i.e. commonly mislabeling one as

another).

ABLE: An Automated Bacterial Load Estimator for the Urinoculture Screening

579

the daily practice of an analysis laboratory. To this

aim, currently, ABLE is being extensively (and suc-

cessfully) tested in DIESSE research laboratories, in

order to compare its responses with those of a team of

expert biologists, who are expected to evidence pos-

sible weaknesses to be solved before its ﬁnal release.

5 CONCLUSIONS

Urinary tract infections can be caused by diverse mi-

crobes, including fungi, viruses, and bacteria. Bacte-

ria are actually the most common cause of UTIs. Nor-

mally, bacteria that enter the urinary tract are rapidly

removed by the body before they cause symptoms.

However, sometimes bacteria overcome the bodys’

natural defenses and, actually, roughly 150 millions

of infections occur annually worldwide. In this pa-

per, an automatic tool, called ABLE, to detect UTIs

and to establish their severity, was described. The

system shows a good accuracy in ﬁnding typical mi-

croorganisms present in humans, and gives no false

negatives. Moreover, it is capable to reveal contami-

nated plates (where multiple infections are present on

the same dish). Preliminary promising experimental

results have been reported by DIESSE biologists, who

are testing ABLE in their laboratories.

REFERENCES

Agah, A., editor (2014). Artiﬁcial Intelligence in Health-

care. CRC Press.

Ahuja, K. and Tuli, P. (2013). Object recognition by

template matching using correlations and phase an-

gle method. International Journal of Advanced Re-

search in Computer and Communication Engineering,

2(3):1368–1373.

Andreini, P., Bonechi, S., Bianchini, M., Mecocci, A., and

Di Massa, V. (2015). Automatic image classiﬁcation

for the urinoculture screening. In Smart Innovation,

Systems and Technologies, volume Intelligent Deci-

sion Technologies, 39, pages 31–42.

Ballabio, C., Venturi, N., Scala, M. R., Mocarelli, P., and

Brambilla, P. (2010). Evaluation of an automated

method for urinoculture screening. Microbiologia Me-

dica, 5(3):178–180.

Bandinelli, N., Bianchini, M., and Scarselli, F. (2012).

Learning long–term dependencies using layered graph

neural networks. In Proceedings of IJCNN–WCCI

2012, pages 1–8.

Belazzi, R., Diomidous, M., Sarkar, I. N., Takabayashi, K.,

Ziegler, A., McCray, A. T., and Sim, I. (2011). Data

analysis and data mining: Current issues in biomedi-

cal informatics support systems. Methods Inf. Med.,

50(6):536–544.

Berlin, A., Sorani, M., and Sim, I. (2006). A taxonomic de-

scription of computer–based clinical decision support

systems. J. Biomedical Informatics, 39:657–667.

Bianchini, M., Maggini, M., and Jain, L. C., editors (2013).

Handbook on Neural Information Processing, volume

Intelligent Systems Reference Library, 49. Springer.

Bourbeau, P. P. and Ledeboer, N. A. (2013). Automation in

clinical microbiology. Journal of Clinical Microbiol-

ogy, 51(6):1658–1665.

Broerm, M. A., Bahçeci, S., Vader, H. L., and Arents, N. L.

(2011). Screening for urinary tract infections with the

sysmex uf–1000i, urine ﬂow cytometer. Journal of

Clinical Microbiology, 49:1025–1029.

Brugger, S. D., Baumberger, C., Jost, M., Jenni, W., Brug-

ger, U., and Múhlemann, K. (2012). Automated

counting of bacterial colony forming units on Agar

plates. PLoS ONE, 7(3):e33695.

Chen, W.-B. and Zhang, C. (2009). An automated bacterial

colony counting and classiﬁcation system. Inf. Syst.

Front., 11(4):349–368.

Clarke, M. L., Burton, R. L., Hill, A. N., Litorja, M.,

Nahm, M. H., and Hwang, J. (2010). Low–cost, high–

throughput, automated counting of bacterial colonies.

Cytometry Part A, 77(8):790–797.

Deserno, T. M., editor (2011). Biomedical Image Process-

ing. Springer–Verlag, New York.

Dey, D. K., Ghosh, S., and Mallick, B. K. (2010). Bayesian

Modeling in Bioinformatics. CRC Press.

Gonzalez, R. and Woods, R. (2008). Digital Image Process-

ing. Addison Wesley.

Heckerling, P. S., Canaris, G. J., Flach, S. D., Tape, T. G.,

Wigton, R. S., and Gerber, B. S. (2007). Predictors

of urinary tract infection based on artiﬁcial neural net-

works and genetic algorithms. Int. J. Med. Inform.,

76(4):289–296.

Henze, N. and Zirkler, B. (1990). A class of invari-

ant consistent tests for multivariate normality. Com-

munications in Statistics – Theory and Methods,

19(10):3595–3617.

Mardia, K. V. (1970). Measures of multivariate skew-

ness and kurtosis with applications. Biometrika,

57(3):519–530.

NHS Purchasing and Supply Agency (2011). Automated

urine screening systems.

Ogawa, H., Nasu, S., Takeshige, M., Funabashi, H., Saito,

M., and Matsuoka, H. (2012). Noise–free accurate

count of microbial colonies by time–lapse shadow im-

age analysis. Journal of Microbiological Methods,

91(43):420–428.

Otsu, N. (1979). A threshold selection method from gray–

level histograms. IEEE Trans. Sys. Man Cyber., 9:62–

66.

Rice, F. and Baruch, A. (2009). Evaluation of BioMérieuxs

PREVI Isola, an automated microbiology specimen

processor: Improving efﬁciency and quality of results.

Torres, A. and Nieto, J. J. (2006). Fuzzy logic in medicine

and bioinformatics. J. of Biomedicine and Biotechnol-

ogy, (91908).

Wang, W. (2011). Colony image acquisition system

and segmentation algorithms. Optical Engineering,

50(12):123001–123010.

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

580