Detection System of Gram Types for Bacteria

from Gram Stained Smears Images

Ryosuke Iida

, Kazuki Hashimoto

, Kouich Hirata

, Kimiko Matsuoka

and Shigeki Yokoyama

Kyushu Institute of Technology, Kawazu 680-4, Iizuka 820-8502, Japan

Osaka General Medical Center, Bandaihigashi 3-1-56, Sumiyoshi, Ohsaka 558-8558, Japan

KD-ICONS, Ohmoriminami 4-6-15-304, Ohta, Tokyo 143-0013, Japan

cby41060@pop01.odn.ne.jp, yokoyama@kd-icons.co,jp

Keywords:

Gram Stain, Gram Stained Smears Images, Gram Types, Gram Positive Cocci, Gram Positive Bacilli, Gram

Negative Cocci, Gram Negative Bacilli.

Abstract:

In this paper, we develop the detection system of Gram types determined by stained colors and stained shapes

for bacteria from Gram stained smears images. Here, we call four types of bacteria, that is, Gram positive

cocci (GPC), Gram positive bacilli (GPB), Gram negative cocci (GNC) and Gram negative bacilli (GPB)

Gram types, and then add to two types as Gram positive unknown (GPU), and Gram positive unknown (GNU).

The system ﬁrst infers the candidate regions of bacteria by using image processing. Next, it constructs a

classiﬁer dividing the candidate regions into Gram types by using SVM (support vetcor machine) and DNN

(deep neural network). Finally, it detects the occurrences of Gram types in a newly input image and retrieves

Gram stained smears images similar as the input image such that the occurrence ratio for the Gram types is

similar.

1 INTRODUCTION

The Gram stain (Bartholomew and Mittwer, 1952) is

the method for microbial smears test in microscope

test, introduced by Hans Christian Gram (1853–1938)

at 1884. For the Gram stain, based on the stained col-

ors as purple/violet or red/pink, the stained shapes as

sphere-shape, rod-shape, singles, pairs, chains, clus-

ters, and so on, we detect bacteria occurring in the

smears for the samples of blood, sputum, feces, pus

and urine.

After Gram staining, we call the bacteria colored

by purple or violet Gram positive and those by red or

pink Gram negative. Also we call the bacteria stained

as sphere-shape cocci and those as rod-shape basilli.

Hence, in this paper, we call four kinds of bacteria

as Gram positive cocci (GPC), Gram positive bacilli

(GPB), Gram negative cocci (GNC) and Gram nega-

tive bacilli (GNB)

Gram types.

Since the Gram stain is applicable inexpensively

and fast returns the results (within 30 min.), it is im-

portant for the initial medical care of infectious dis-

eases (Mitsuda, 2004; Yamamoto, 2015). On the

Sometimes we call GPB and GNB Gram positive rod

and Gram negative rod (Smith et al., 2018).

other hand, Gram staining is possible to stain not

only bacteria but also non-bacteria substances such

as leukocytes, dust, oil and crystals. Also, there ex-

ist many kinds of phlogogenic fungus for infectious

diseases. Table 1 illustrates the relationship between

Gram types and bacteria.

Table 1 shows that the bacteria as the phlogogenic

fungus for hospital-acquired infection tend to belong

to GPC or GNB. Then, the detected bacteria will

determine the direction for culture and identiﬁcation

tests (Mitsuda, 2004; Yamamoto, 2015).

In the microscope test, Gram stained smears im-

ages are checked manually and visually and not au-

tomatically in general. The reason is that we can de-

tect bacteria exactly by applying culture and identi-

ﬁcation tests after the microscope test. On the other

hand, anaerobic bacteria are never lived in the culture

test, they cannot be detected by the identiﬁcation test.

Hence, the detected bacteria through the culture and

identiﬁcation tests are the part of bacteria in smears.

Also, since the culture and identiﬁcation tests

spend one day, we cannot apply them to the initial

medical care of infectious diseases. Furthermore,

whereas expert skills are necessary to detect bacte-

ria manually and visually from Gram stained smears

Iida, R., Hashimoto, K., Hirata, K., Matsuoka, K. and Yokoyama, S.

Detection System of Gram Types for Bacteria from Gram Stained Smears Images.

DOI: 10.5220/0008964404770484

In Proceedings of the 9th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2020), pages 477-484

ISBN: 978-989-758-397-1; ISSN: 2184-4313

477

Table 1: The relationship between Gram stain and bacteria.

Gram types bacteria

(color, shape)

Gram positive cocci Enterococcus faecalis,

(purple/violet, sphere) Staphylococcus aureus,

Streptcoccus pyogenes,

Streptcoccus pneumoniae

Gram positive bacilli Clostridium

(purple/violet, rod) (Clostridium perfringens,

Clostridium tetani),

Corynebacterium diphtheriae,

Listeria monocytogenes

Gram negative cocci Nesseria meningitidis,

(red/pink, sphere) Nesseria gonorrhoeae

Gram negative bacilli Bacteroides spp.,

(red/pink, rod) Klebsiella spp.,

Psuedomonas aeruginosa,

Escherichia coli, Serratia spp.

Gram intermediate Legionella spp.,

Mycobacterium spp.

images, such technicians with expert skills are not

enough to apply the initial medical care in Japan.

Hence, the automatic detection of bacteria from Gram

stained smears images automatically is required.

In order to solve these problems, in this paper, as

the prepossessing of detecting bacteria exactly from

Gram stained smears images, we focus on the Gram

types. Then, we develop the system to detect such

types of bacteria and to compute the occurrence ratio

of types in every image.

In this system, ﬁrst we extract candidate regions of

bacteria as the regions obtained by excluding the re-

gion not occurring bacteria from Gram stained smears

images by using image processing. Then, we compute

the features as colors, areas, aspect ratio, and so on.

After classifying candidate regions by colors as Gram

positive and Gram negative, we detect Gram types by

using the classiﬁer constructed from SVM (support

vector machine) and DNN (deep neural network).

1.1 Related Works

As related works to this paper, Carvajal et al. (Car-

vajal et al., 2014) have developed the system to learn

the candidate areas from ﬁxed-size (51 × 38 pixels)

images applicable to the microscope test with high

magniﬁcation. They have dealt with the Gram stained

smear images per 64× ﬁeld. Note that we use images

par 1, 000× ﬁeld in the microscope test in general,

which we deal with in this paper. Then, whereas they

have dealt with Gram stained smear images, they have

not achieved to detect Gram types or bacteria.

On the other hand, Smith et al. (Smith et al., 2018)

have dealt with the Gram stained smear images pro-

vided from the sample of blood. Their target Gram

types are GNB, GPC in clusters and GPC in pairs or

chains, which are meaningful Gram types for blood.

Then, after extracting ﬁxed size (146 × 146 pixels)

images, they have detected the above three Gram

types by using CNN.

Note that the above researches of (Carvajal et al.,

2014) and (Smith et al., 2018) have dealt with ran-

domly selected ﬁxed size images as training data, so

their researches have not detect the area of bacte-

ria in the whole images. On the other hand, Lejon

and Andersson (Lejon and Andersson, 2016) have de-

tected the area of bacteria in the image, as same as

this paper, and then Gram types and then bacteria

from the Gram stained smear images for the sample

of blood by MATLAB. Whereas they have detected

Gram types and bacteria without machine learning,

we detect Gram types by using machine learning.

In (Lejon and Andersson, 2016), they have

adopted the template matching to detect the areas of

bacteria, which they have implicitly adopted the ideal

assumption that every bacterium has the similar small

size and there exist no substances such as dust with

the similar size of bacteria. Furthermore, when deal-

ing with the images for not only blood but also spu-

tum and feces, more kinds of bacteria such as Table 1

containing the bacteria not occurring in the images for

blood and many other substances except bacteria oc-

cur in the images for sputum and feces.

Hence, we can position this paper to develop a

new system to detect the area of bacteria and then

Gram types with machine learning applicable to the

Gram stained smear images for not only the sample

of blood but also other samples of sputum, feces, pus

and urine uniformly.

2 DETECTING SYSTEM OF

GRAM TYPES

In this section, we explain our detection system of

Gram types from Gram stained smears images.

2.1 Data and Outline of System

First, we use the Gram stained smears images per

1, 000× ﬁeld, provided from Osaka General Medical

Center applied to our detecting system. Every image

consists of 2, 448 × 1, 920 pixels and is assigned the

ratio of the occurrences of GPC, GPB, GNC and GNB

as percentage. For every sample, the number of im-

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

478

ages is 42 for blood, 40 for sputum, 10 for feces, 40

for pus and 69 for urine, respectively.

Our detection system of Gram types mainly con-

sists of the extraction of candidate regions of bac-

teria, learning phase and detecting and retrieving

phase. First, by using image processing, the sys-

tem extracts candidate regions of bacteria from Gram

stained smears images. Then, in the learning phase,

the system constructs the classiﬁer for Gram types

from training data. Finally, in the detecting and

retrieving phase, the system detects the regions of

Gram types by using the classiﬁer, and then outputs

the input image with depicted such regions and re-

trieves similar images such that the occurrence ratios

of Gram types are similar.

2.2 Candidate Regions of Bacteria

In order to extract candidate regions of bacteria, in our

system, ﬁrst we convert Gram stained smears images

to glayscale images. Here, we adopt the NTSC (non-

subsampled contoulet transform) coefﬁcient method

supported from OpenCV

and then compute lumi-

nance.

Next, we transform the grayscale images to the bi-

nary images by applying binarization consisting of 1

if a pixel has the luminance more than the threshold

and 0 otherwise. Here, in our binarization, we adopt

adaptive thresholding (Keahler and Bradski, 2013)

to determine the threshold by considering pixels on

neighbors at a current pixel. As a result, our sys-

tem extracts the regions with clearer boundaries than

around pixels. Here, the adaptivethresholding is com-

puted by matrix and the coefﬁcient of neighbor re-

gions under Gaussian distribution minus a subtractive

constant

. The constant reﬂects ﬂuctuation such that

noises tend to be ignored if the constant is large and

pixels with small changes of luminance tend to be re-

mained if it is small.

Then, by using morphorogical operation (Keahler

and Bradski, 2013), we separate connected pixels

from other pixels and then eliminate noises. Here, we

adopt opening processing as applying the dilation at

n times after applying the erotions at n times.

Finally, we extract edges as boundaries as pixels

with large changes of luminances. We call the set of

edges for a substance an outline. In our system, we

adopt a Canny ﬁlter (Canny, 1986) as a robust edge

detection ﬁlter to noises. By applying the edge de-

https://docs.opencv.org/3.1.0/de/d25/

imgproc color conversions.html

https://docs.opencv.org/3.0-beta/modules

/imgproc/doc/miscellaneous

transformations.html

?highlight=cv2.adaptive#cv2.adaptiveThreshold

tection ﬁlter, we exclude regions as noises with small

changes of luminances.

After determining the regions consisting of sub-

stances but not noises, we regard the regions such that

every value of areas, aspect ratio and circularity satis-

ﬁes every threshold as bacterial regions. As a result,

we extract such bacterial regions as candidate regions

of bacteria.

2.3 Training Data

In order to construct training data, we design the tool

as Figure 1 (whose GUI is in Japanese) to represent

the rectangle as candidate regions of bacteria, and

then assign the correct Gram type to the rectangle

manually through GUI by the fourth author who is the

medical technologist for clinical microbial testing.

Figure 1: Tool for constructing training data.

As a result, Table 2 illustrates the number of

training data as images transferring a single bac-

terium. Here, since there exist regions not determin-

ing whether cocci or bacilli, we add those as Gram

positive unknown (GPU) and Gram negative unknown

(GNU) and extend Gram types to six kinds of GPC,

GPB, GPU, GNC, GNB and GNU.

Table 2: The number of training data for samples.

sample num. color cocci bacilli unknown total

blood 6,216

pos. 1,062 512 406 1,980

neg. 0 3,424 812 4,236

sputum 10,308

pos. 1,118 665 1,821 3,604

neg. 826 2,522 3,356 6,704

feces 2,905

pos. 20 74 154 248

neg. 56 2,029 572 2,657

pus 8,178

pos. 296 534 1,388 2,218

neg. 0 4,225 1,735 5,960

urine 6,052

pos. 14 50 395 459

neg. 0 4,397 1,196 5,593

Detection System of Gram Types for Bacteria from Gram Stained Smears Images

479

2.4 Learning Phase

From training data illustrated as Table 2, in the learn-

ing phase, we construct the classiﬁers by using SVM

(support vector machine) and DNN (deep neural net-

work). Here, we adopt SVM as the library pro-

vided from OpenCV (Keahler and Bradski, 2013)

whose kernel is CHI2 (Li et al., 2010). Also we

adopt DNN as Caffe (convolutional architecture for

fast feature embedding) based on the rayer structure

of AlexNet (Krizhevsky et al., 2017).

For SVM, we construct the classiﬁer to detect

Gram types by using feature values as features for

candidate regions of bacteria. Here, we adopt features

as the area, the aspect size, the aspect ratio, the color

and the circularity of the region, and also the number

of detected bacteria in the region.

On the other hand, for DNN, we construct the

classiﬁer to detect Gram types by using images for

candidate regions of bacteria as training data.

2.5 Detecting and Retrieving Phase

In the detecting and retrieving phase, after setting a

sample and a ﬁle path, referring to the button “Re-

fer,” on GUI, our system starts to detect four Gram

types. Then, it outputs, for each of SVM and DNN,

the occurrence ratio of every Gram type, the main im-

age such that the substance is enclosed by a colored

rectangle if it belongs to one Gram type and the three

subimages whose occurrence ratio is similar. Here,

we can change the main image or subimages by SVM

and DNN.

Figure 2 illustrates the output of our detection sys-

tem for the sample of blood and DNN. Here, in the

main image, the Gram types of GPC, GPB, GNC,

GNB, GPU and GNU are enclosed by frames colored

by blue, light blue, red, pink, light green and black,

respectively.

In Figure 2, the left windows in the detection sys-

tem represents the occurrence ratio for every Gram

type. The upper three images are the original image,

the result by SVM and the result by DNN from left to

right. The right three images are similar images as the

original image.

3 EXPERIMENTAL RESULTS

In this section, we give experimental results for our

detecting system of Gram types. Here, computer

environment to the learning phase is CPU Intel(R)

Xeon(R) E5-1603 v3 @2.80GHz, RAM 32.0GB and

OS Windows 10 Pro for Workstations. The samples

consist of blood, sputum, feces, pus and urine. In this

section, we use the images to assigned the occurrence

ratio of Gram types by the professional technician,

which we call the images to assigned ratios. Then,

the number of the images to assigned ratios for blood,

sputum, feces, pus and urine is 42, 40, 10, 40 and 69,

respectively, and the total number of the images to as-

signed ratios is 201.

3.1 Parameters

Table 3 illustrates the parameters for image process-

ing for Section 2.2 applied to experimental results.

Here, we denote the size of neighbor for adaptive

thresholding by N, the subtractive constant by C and

the number of opening processings by P. Also we de-

note the upperbound an the lowerbound for the area

as the candidate regions of bacteria by S

min

and S

max

those for the aspect ratio by A

min

and A

max

, and those

for the circularity by C

min

and C

max

Table 3: The parameters for image processing applied to

experimental results.

sample N C P S

min

max

min

max

min

max

blood 257 37 2 200 30000 0.1 1.0 0.1 1.0

sputum 431 37 2 130 30000 0.1 1.0 0.1 1.0

feces 281 18 2 300 30000 0.1 1.0 0.1 1.0

pus 301 38 2 150 15000 0.1 1.0 0.1 1.0

urine 581 35 2 200 50000 0.1 1.0 0.1 1.0

3.2 Detection of Gram Classes

Figures 3, 4 and 5 illustrate the results obtained by de-

tecting Gram types by SVM and DNN from arbitrary

selected images for blood, sputum and feces and by

searching their similar images.

The images in Figures 3, 4 and 5 are those to as-

signed ratios. Then, Table 4 illustrates the assigned

and detected occurrence ratios of Gram types for

blood (Figure 3), sputum (Figure 4) and feces (Fig-

ure 5) by SVM and DNN.

Table 4 shows that, in these images, whereas the

detected occurrence ratios by SVM is more similar

as the assigned occurrence ratios than those by DNN,

the accuracy is too insufﬁcient to detect Gram types

exactly. For all the cases, every detected Gram type

contains GPU and GNU. A large ratio of GNU fol-

lows that the stained dusts are detected as GNU.

3.3 Evaluation for Detection

In order to evaluate the method to detect the Gram

types, by using the assigned ratio of the occurrences

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

480

Figure 2: The detection system of Gram types for the sample of blood.

SVM DNN

similar images

Figure 3: The images for blood by detecting Gram types by

SVM and DNN (upper) and its similar images (lower).

of GPC, GPB, GNC and GNB as percentage, we ap-

ply the following test for sputum and pus.

1. First, we divide all the images for a sample into

two groups in half randomly. We call one group

training images and another group test images.

Here, we obtain 21 training images and 21 test

images for blood, 20 training images and 20 test

images for sputum and pus.

2. From the training images, we construct the train-

ing data (regions) with the classes of GPC, GPB,

GPU, GNC, GNB and GNU as same as Sec-

tions 2.2 and 2.3.

SVM DNN

similar images

Figure 4: The images for sputum by detecting Gram types

by SVM and DNN (upper) and its similar images (lower).

3. By using SVM and DNN, we construct classi-

ﬁer from training data and apply it to test im-

ages. Then, for every test image, we compare the

assigned ratio of the occurrences of GPC, GPB,

GNC and GNB to the image with the detected ra-

tio of the occurrences of GPC, GPB, GPU, GNC,

GNB and GNU from the image.

4. Repeat the above procedures at ﬁve times.

In order to evaluate the assigned ratio to the im-

age and the detected ratio from the image, we ﬁrst

compare the occurrences for color. Let p

and p

the assigned ratio and the detected ratio of the occur-

rences of Gram positive bacteria, respectively. Then,

Detection System of Gram Types for Bacteria from Gram Stained Smears Images

481

SVM DNN

similar images

Figure 5: The images for feces by detecting Gram types by

SVM and DNN (upper) and its similar images (lower).

Table 4: The assigned (ass.) and detected occurrence ra-

tios (occ. ratio) of Gram types (GCs) for blood (Figure 3),

sputum (Figure 4) and feces (Figure 5) by SVM and DNN.

blood (Figure 3) sputum (Figure 4)

GCs occ. ratio

ass. SVM DNN

GPC 0 0.00 0.00

GPR 0 0.00 2.74

GNC 0 0.00 17.81

GNR 100 83.56 41.10

GPU – 2.74 0.00

GNU – 13.70 38.36

GCs occ. ratio

ass. SVM DNN

GPC 10 6.92 10.77

GPR 50 33.85 31.54

GNC 0 0.77 10.77

CNR 40 26.15 30.00

GPU – 2.31 0.77

GNU – 30.00 16.15

feces (Figure 5)

GCs occ. ratio

ass. SVM DNN

GPC 0 0.00 0.00

GPR 10 6.25 6.25

GNC 0 0.00 13.75

GNR 90 70.00 46.25

GPU – 8.75 8.75

GNU – 15.00 25.00

we say that a test image has an admissible color if

− p

| ≤ 20%.

Table 5 illustrates the number of images with an

admissible color in the test images for blood, sputum

and pus. Here, the number of the test images for blood

is 21 and that for sputum and pus is 20.

Table 5 shows that the ratio for the test images

with an admissible color is more than 50% and, in

particular, it is about 90% for pus by DNN.

Next, we compare the occurrences for shapes. Let

and c

be the assigned ratio and the detected ratio

of the occurrences of cossi, respectively. Also let b

and b

be the assigned ratio and the detected ratio of

the occurrences of basilli, respectively. Then, we say

that a test image has admissible cossi if |c

− c

| ≤

Table 5: The number of images with an admissible color in

the test images for blood, sputum and pus.

blood, SVM (21) blood, DNN (21)

1 2 3 4 5 ave.

12 10 14 14 12 12.4

1 2 3 4 5 ave.

12 12 13 13 13 12.6

sputum, SVM (20) sputum, DNN (20)

1 2 3 4 5 ave.

14 11 9 8 11 10.6

1 2 3 4 5 ave.

9 12 12 11 9 10.6

pus, SVN (20) pus, DNN (20)

1 2 3 4 5 ave.

9 12 12 11 9 10.6

1 2 3 4 5 ave.

18 18 17 17 18 17.6

20% and admissible basilli if |b

− b

| ≤ 20%. Fur-

thermore, we say that a test image has admissible

shapes if |c

− c

| ≤ 20% and |b

− b

| ≤ 20%.

Table 6 illustrates the number of images with ad-

missible cossi, basilli and shapes in the test images for

sputum and pus. Here, the number of the test images

is 20, respectively. Also we do not count the number

of GPU and GNU.

Table 6: The number of images with admissible cossi (C),

basilli (B) and shapes (S) in the test images for blood, spu-

tum and pus.

blood, SVM (21) blood, DNN (21)

1 2 3 4 5 ave.

C 18 13 16 18 16 16.2

B 12 13 12 10 14 12.2

S 9 6 7 8 9 7.8

1 2 3 4 5 ave.

C 18 5 18 10 15 13.2

B 16 6 20 11 18 14.2

S 15 5 18 10 15 12.6

sputum, SVM (20) sputum, DNN (20)

1 2 3 4 5 ave.

C 7 9 6 5 6 6.6

B 8 7 9 10 10 8.8

S 0 0 0 0 0 0

1 2 3 4 5 ave.

C 8 5 7 5 5 6

B 8 13 8 7 8 8.8

S 5 5 5 5 5 5

pus, SVN (20) pus, DNN (20)

1 2 3 4 5 ave.

C 6 6 6 6 5 5.8

B 9 13 7 7 8 8.8

S 4 6 4 6 5 5

1 2 3 4 5 ave.

C 4 5 7 6 9 6.2

B 5 2 4 9 3 4.6

S 2 0 2 4 3 2.2

Table 6 shows that, for blood, the accuracy of de-

tecting shapes (cossi and basilli) tends to be higher

than the accuracy of detecting colors and the ratio

for the test images with admissible cossi and basilli

is greater than 50%. Since the second round for

DNN detects all the assigned labels of GNB to GNC,

the number of the images with admissible cossi and

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

482

basilli is very small.

On the other hand, for sputum and pus, the ac-

curacy of detecting shapes (cossi and basilli) is much

lower than the accuracy of detecting colors in Table 5.

In all the cases, the ratio for the test images with ad-

missible cossi, basilli and shapes for sputum and pus

is less than 50%.

In particular, for sputum by SVM, the reason why

the number of test images with admissible shapes is

0 is that the detection of cossi (resp., basilli) suc-

ceeds when the ratio of the occurrencesof cossi (resp.,

basilli) is near to 0%. Also, for pus by SVM, either

the detected ratio of the occurrences of GNB is 0%

and that of GNC is 100% or the detected ratio of the

occurrences of GNC is 0% and that of GNB is 100%.

For pus by DNN, the detected ratio of the occurrences

of GNB is always 0% and GNB is determined as GNC

in the detection.

Hence, it is necessary to improve our detection

system of detecting shapes rather than color for spu-

tum and pus.

3.4 Running Time

Finally, Table 7 illustrates the average running time to

detect the Gram type for a single candidate region of

bacteria by SVM and DNN for every sample.

Table 7: The average running time (ms) to detect the Gram

types for a single candidate region of bacteria by SVM and

DNN for every sample.

SVM DNN

blood 0.033 240.509

sputum 0.143 247.572

feces 0.030 242.059

pus 0.081 246.837

urine 0.039 245.090

Table 7 shows that the running time to detect by

SVM is much faster than that by DNN.

4 CONCLUSION

In this paper, we have developed the detecting system

of Gram types for bacteria from Gram stained smears

images. By applying our system to 201 Gram stained

smears images to assigned the occurrence ratios of

Gram types, we have given the experimental results

for our system.

Since our system is still proto-typing and there ex-

ist many future works to improve our system. First of

all, we have just applied standard image processing

to our system and then not designed the method ap-

propriate to Gram stained smears images, so it is an

important future work to design such a method and

embed it to our system.

As stated in Section 3.3, it is necessary to improve

the detection of shapes rather than color for sputum

and pus. Then, it is a future work to apply the image

processing methods such as black top-hat transform,

label connected component and template matching

proposed by (Lejon and Andersson, 2016) and then to

analyze which of them is useful of our system. Also,

since the labels consist of GPC, GPB, GNC and GNB,

it is a future work to design the method to evaluate the

detection.

Furthermore, our system cannot avoid to detect

dust as bacteria completely illustrated in Figure 6

(left), so it is a future work to improve our system

to avoid to this situation. Also it is necessary to detect

leukocyte phagocystosis in Figure 6 (right), which is

an important future work.

Figure 6: The detection of dusts as bacteria (left) and leuko-

cyte phagocystosis (right).

ACKNOWLEDGMENTS

This work is partially supported by Grant-in-Aid

for Scientiﬁc Research 17H00762, 16H02870 and

16H01743 from the Ministry of Education, Cul-

ture, Sports, Science and Technology, Japan and the

next generation innovation project 2020 from Tokyo

Metropolitan Small and Medium Enterprise Support

Center.

REFERENCES

Bartholomew, J. and Mittwer, T. (1952). The Gram stain.

Bacteriol. Rev., 16:1–29.

Canny, J. (1986). A computional approach to edge detec-

tion. IEEE Trans. Patt. Anal. Mach. Intel., 8:679–698.

Carvajal, J., Smith, D., Zhao, K., Wiliem, A., Finucane, P.,

Hobson, P., Jennings, A., McDougall, R., and Lovell,

B. (2014). An early experience toward developing

computer aided diagnosis for Gram-stained smear im-

ages. In Proc. CVPR’14, pages 62–28.

Detection System of Gram Types for Bacteria from Gram Stained Smears Images

483

Keahler, A. and Bradski, G. (2013). Learning OpenCV:

Computer vision in C++ with the OpenCV library.

O’Reilly Media.

Krizhevsky, A., Sutskever, I., and Hinton, G. (2017). Im-

ageNet classiﬁcation with deep convolutional neural

network. Comm. ACM, 60:84–90.

Lejon, S. and Andersson, E. (2016). Semi-automatic

segmentation, detection and classiﬁcation of Gram

stained bacteria in blood sample. Master Thesis,

Lund University.

Li, F., Carreira, J., and Sminchisescu, C. (2010). Object

recognition as ranking holistic ﬁgure-ground hypothe-

ses. In Proc. CVPR’10, pages 1712–1719.

Mitsuda, T. (2004). Foundations of clinical microbial test-

ing for medical care of infectious diseases and infec-

tion control (in Japanese). International Medical Pub-

lisher.

Smith, K., Kang, A., and Kirby, J. (2018). Automated in-

terpretation of blood culture Gram stains by use of a

deep convolutional neural network. J. Clin. Microbio.,

56:e01521–17.

Yamamoto, T. (2015). Medical care support for infectious

disceases by using Gram stain (in Japanese). J. Japan.

Soc. Cli. Micribio., 25:265–276.

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

484