Object Detection as Campylobacter Bacteria and Phagocytotic Activity

of Leukocytes in Gram Stained Smears Images

Kyohei Yoshihara and Kouich Hirata

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

Keywords:

Object Detection, Campylobacter Bacteria, Phagocytotic Activity of Leukocytes, Gram Stained Smears

Images, Faster R-CNN, RetinaNet, YOLOv5.

Abstract:

In this paper, we apply object detection to Gram stained smear images, where objects are Campylobacter

bacteria and phagocytotic activity of leukocytes. Then, we adopt three CNN-based object detectors of Faster

R-CNN, RetinaNet and YOLOv5. The outline of the detection is ﬁrst to annotate the regions of objects as

Campylobacter bacteria and phagocytotic activity of leukocytes in training images, and then to detect the

regions of objects in the remained test images by using the detectors. Finally, we give experimental results of

detecting Campylobacter bacteria and phagocytotic activity of leukocytes in Gram stained smear images by

using the detectors.

1 INTRODUCTION

The Gram stain (Bartholomew and Mittwer, 1952) is

the method for microbial smears test in microscope

test per 1, 000× ﬁeld, introduced by Hans Christian

Gram (1853–1938)at 1884. For the Gram stain, based

on the stained colors as purple/violet or red/pink, the

stained shapes as sphere-shape, rod-shape, singles,

pairs, chains, clusters, and so on, we detect bacteria

occurring in the smears for the samples of blood, spu-

tum, 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, we can classify bacteria into the four kinds

as Gram positive cocci, Gram positive bacilli, Gram

negative cocci and Gram negative bacilli

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. On the other hand, Gram staining is possible

to stain not only bacteria but also non-bacteria sub-

stances such as leukocytes, dusts, oil and crystals.

Also there exist many kinds of phlogogenic bacteria

for infectious diseases.

In the microscope test, Gram stained smears im-

ages are checked manually and visually but not auto-

Sometimes we call basilli rods (Smith et al., 2018).

matically in general. The reason is that we can detect

bacteria exactly by applying the culture test and the

identiﬁcation test after the microscope test. On the

other hand, since anaerobic bacteria are never lived in

the culture test, they cannot be detected by the iden-

tiﬁcation test. Then, the detected bacteria through the

culture test and the identiﬁcation test are the part of

bacteria in smears.

Also since the culture test and the identiﬁcation

test spend one day, we cannot apply them to the ini-

tial medical care of infectious diseases. Furthermore,

whereas expert skills are necessary to detect bacte-

ria manually and visually from Gram stained smears

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 is required.

In this paper, we focus on Campylobacterbacteria

and phagocytotic activity of leukocytes. The Campy-

lobacter bacteria are Gram negative basilli and phlo-

gogenic bacteria causing so called Campylobacter en-

teritis. Also the phagocytotic activity of leukocytes

works as the natural immunity. Then, the detection of

them is an important tasks for the microscope test of

the Gram stained smears images. However, it is also

well-known to be difﬁcult in the microscope test.

The difﬁculty of the detection of Campylobac-

ter bacteria is that we cannot distinguish the Campy-

lobacter bacteria from dusts, since the Campylobacter

bacteria are as small as dusts and their shapes are also

534

Yoshihara, K. and Hirata, K.

Object Detection as Campylobacter Bacteria and Phagocytotic Activity of Leukocytes in Gram Stained Smears Images.

DOI: 10.5220/0010813800003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 534-541

ISBN: 978-989-758-549-4; ISSN: 2184-4313

similar as dusts. Figure 1 illustrates the image con-

taining Campylobacter bacteria, which occur in just

two red circles, and a Campylobacter bacterium in

Gram stained smears images.

Figure 1: The Gram stained smears image containing

Campylobacter bacteria (upper) and a Campylobacter bac-

terium in Gram stained smears images (lower).

The difﬁculty of the detection of phagocytotic ac-

tivity of leukocytes is that not only phagocytotic im-

ages that a leukocyte enclose bacteria as avoiding to

its nucleus but also quasi-phagocytotic images that a

leukocyte and bacteria are just overlapping are ob-

served in Gram stained smears images. Figure 2 il-

lustrates a phagocytotic image, a quasi-phagocytotic

image and a non-phagocytotic image in Gram stained

smears images.

phagocytotic quasi-phagocytotic non-phagocytotic

Figure 2: A phagocytotic image, a quasi-phagocytotic im-

age and a non-phagocytotic image in Gram stained smears

images.

Recently, Yoshihara and Hirata (Yoshihara and

Hirata, 2021) have classiﬁed Campylobacter bacte-

ria and phagocytotic activity of leukocytes, respec-

tively, by using VGG16 (Simonyan and Zisserman,

2015) and its improvement with high accuracy. In

their work, for Campylobacter bacteria, after manu-

ally extracting the regions in which both a Campy-

lobacter bacterium occurs and no Campylobacter bac-

terium occurs, they have classiﬁed them. Also, for

phagocytotic activity of leukocytes, after manually

extracting the regions as phagocytotic images, quasi-

phagocytotic and non-phagocytoticimages, they have

classiﬁed them. Hence, the purpose of their work is

to classify Campylobacter images and phagocytotic

images which have been manually extracted, without

object detection.

On the other hand, in this paper, we detect them

with object detection. For Campylobacter bacte-

ria, after annotating the regions of Campylobacter

bacteria in training images, we detect the regions

of Campylobacter bacteria in the remained test im-

ages. For phagocytotic activity of leukocytes, after

annotating phagocytotic, quasi-phagocytoticand non-

phagocytoticregions in training images, we detect the

regions of phagocytotic activity of leukocytes in the

remained test images.

In this paper, we adopt three CNN-based object

detectors of Faster R-CNN (Ren et al., 2015), Reti-

naNet (Lin et al., 2017) and YOLOv5 (Jocher, 2020).

By using these detectors, we give experimentalresults

of detecting Campylobacter bacteria and phagocytotic

activity of leukocytes in Gram stained smears images.

1.1 Related Works

As the works dealing with Gram stained smears im-

ages, Carvajal et al. (Carvajal et al., 2014) have de-

veloped the system to learn the candidate areas from

ﬁxed-size (51 × 38 pixels) images applicable to the

microscope test with high magniﬁcation. Hashimoto

et al. (Hashimoto et al., 2020) have developed the

system to detect Geckler classiﬁcation deﬁned by

the number of buccal squamous epithelial cells and

leukocytes for the Gram stained smears images per

100× ﬁeld for the sample of sputum, in order to guar-

antee howthe Gram stained smearsimage per 1, 000×

ﬁeld is quality for the microscope testing.

Lejon and Andersson (Lejon and Andersson,

2016) have developed the system to classify the bac-

teria occurring in the areas for the sample of blood

by using the template matching. Smith et al. (Smith

et al., 2018) have classiﬁed Gram negative basilli,

Gram positive cossi in clusters and Gram positive

cossi in pairs or chains from the Gram stained smear

images for the sample of blood by using CNN, after

extracting ﬁxed size (146 × 146 pixels) images. Iida

et al. (Iida et al., 2020) have developed the system to

classify four kinds of Gram positive cocci, Gram pos-

itive bacilli, Gram negative cocci and Gram negative

bacilli from the Gram stained smear images for not

only blood but also other samples by using CNN.

Object Detection as Campylobacter Bacteria and Phagocytotic Activity of Leukocytes in Gram Stained Smears Images

535

2 OBJECT DETECTION

The purpose of this paper is to detect Campylobac-

ter bacteria and phagocytotic activity of leukocytes in

Gram stained smears images. The outline of the de-

tection is ﬁrst to annotate the regions of objects as

Campylobacter bacteria and phagocytotic activity of

leukocytes in training images, and then to detect the

regions of objects in the remained test images by us-

ing the detectors.

2.1 Detectors

In this paper, we adopt the following three detectors.

Here, we implement them through PyTorch, which is

an open source machine learning library for Python.

2.1.1 Faster R-CNN

Faster R-CNN (Ren et al., 2015) is a two-stage de-

tector and consists of RPN (region proposal network)

and RoI (region of interest) pooling layer. Then,

it classiﬁes objects by the RoI pooling layer after

proposing regions by the RPN.

To implement Faster R-CNN, we use Detec-

tron2 (Wu et al., 2019) as an object detection library,

and tune up the model of FRN+ResNeXt-101-32x8d

in PyTorch.

2.1.2 RetinaNet

RetinaNet (Lin et al., 2017) is an one-stage detec-

tor and consists of a feature pyramid network back-

bone on the top of a feedforward ResNet architec-

ture. ResNet consists of subnetworks for classifying

anchor boxes and those for regressing from anchor

boxes to ground truth object boxes. This network de-

sign is intentionally simple, which is a reason why

one-stage detectors are faster than two-stage detec-

tors.

To implement RetinaNet, we use Detectron2 (Wu

et al., 2019) as an object detection library, and tune up

the model of ResNet101 in PyTorch.

2.1.3 YOLOv5

YOLOv5 (Jocher, 2020), where YOLO is an acronym

“You only look once” and v5 means “version 5,” is

an one-stage detector integrating of the entire object

detection and classiﬁcation process in a single net-

work. The network of YOLO has 24 convolutional

layers followed by 2 fully connected layers. Then,

YOLO pertains the convolutional layers on the Im-

ageNet classiﬁcation task at half the resolution and

then double the resolution for detection.

To implement YOLOv5, we use a default PyTorch

library as an object detection library and tune up the

model of YOLOv5x in PyTorch.

2.2 Setting

In this paper, our computer environment is under

Google Colab that OS is Ubuntu 18.04.5 LTS, CPU

is Intel(R) Xeon(R) CPU @ 2.20GHz, RAM is 25GB

and GPU is Tesla P100-PCIE.

Also Table 1 illustrates the values of hyperparam-

eters such as epoch, batch

size and lr (learning rate).

Table 1: The values of hyperparameters that epoch,

batch

size and lr.

detector epoch batch size lr

Faster R-CNN 200 16 0.001

RetinaNet 200 16 0.001

YOLOv5 400 16 0.01

To determine the values of hyperparameters, we

tune up manually them that the values of mAP or

other AP’s are large.

2.3 Annotation

In this paper, we use 67 Gram stained smear images

for the sample of feces to detect Campylobacter bac-

teria and 101 Gram stained smear images for the sam-

ple of sputum to detect phagocytotic activity of leuko-

cytes.

Then, with helping the opinions of the medical

technologist, we annotate the regions of object as

Campylobacter bacteria and phagocytotic activity of

leukocytes in training images. For Campylobacter

bacteria, we annotate the region where an Campy-

lobacter bacteria occurs. On the other hand, for

phagocytotic activity of leukocytes, we annotate the

regions where leukocytes have phagocytotic activity

(phagocytotic), have the look of phagocytotic activ-

ity but not (quasi-phagocytotic) and have no phago-

cytotic activity (non-phagocytotic).

Figure 3 illustrates the images of annotating Cam-

phylobacter bacteria and leukocytes.

In the detection of Campylobacter bacteria (resp.,

phagocytotic activity of leukocytes), we use about

85% (resp., 90%) of images as training images includ-

ing validation images and the remained about 15%

(resp., 10%) images as test images. Since the num-

ber of training images is too small, we increase them

at triple by applying data augmentation. After anno-

tating, we resize 640×640 pixels for all the images.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

536

Campylobacter bacteria leukocytes

Figure 3: Annotations for Campylobacter bacteria (left) and

leukocytes (right).

2.4 Evaluation

In order to evaluate the results of the detection,

we adopt the standard measures (Everingham et al.,

2010) for object detection. First, we introduce the fol-

lowing intersection over union (IoU) between the area

P of the predicted box and the area T of the ground

truth box:

IoU =

P∩ T

P∪ T

For a given threshold δ (%), let TP be the number of

the predicted boxes such that IoU ≥ δ, FP the number

of the predicted boxes such that IoU < δ and FN the

number of the ground truth boxes such that IoU < δ.

Then, the standard measures of precision and recall

are deﬁned as follows.

precision =

TP+ FP

, recall =

TP+ FN

Also, an average precision for δ (APδ) is deﬁned as

the averagedetection precision under different recalls.

We use AP when δ = 50 and δ = 75, that is, AP50

and AP75. Furthermore, we adopt a (COCO) mean

AP (mAP) that is an average of APs when varying δ

is from 50 to 95 with a step of 5.

3 EXPERIMENTAL RESULTS

In this section, we give the experimental results to de-

tect Camylobacter bacteria and phagocytotic activity

of leukocytes.

3.1 Detection of Campylobacter

Bacteria

For the detection of Campylobacter bacteria, we use

57 training images including 10 validation images and

Figure 4: The Gram stained smear image such that Campy-

lobacter bacteria are annotated correctly.

10 test images for total 67 images. Then, Table 2 il-

lustrates the values of mAP, AP50 and AP75 to detect

Camphylobacter bacteria by using three detectors.

Table 2: The values of mAP, AP50 and AP75 to detect Cam-

phylobacter bacteria (%).

detector mAP AP50 AP75

Faster R-CNN 5.7 17.9 1.6

RetinaNet 4.1 10.5 1.6

YOLOv5 14.6 43.6 4.1

Table 2 shows that YOLOv5 has larger values of

mAP, AP50 and AP75 than Faster R-CNN and Reti-

naNet. On the other hand, for all of the Faster R-

CNN, RetinaNet and YOLOv5, the values of AP75 is

much smaller than those of mAP and AP50.

By the deﬁnition of mAP, the threshold δ such that

the value of APδ is equal to that of mAP is in the range

of [50, 75]. Then, for the detected regions of Campy-

lobacter bacteria, the overlap between the predicted

box and the ground truth box is not large as IoU.

Next, we represent the results of detecting Campy-

lobacter bacteria in Gram stained smear images. Fig-

ure 4 illustrates the Gram stained smear image such

that Campylobacter bacteria are annotated correctly.

Then, Figure 5 illustrates the result of detect-

ing Camphylobacter bacteria by Faster R-CNN, Reti-

naNet and YOLOv5 from the Gram stained smear im-

ages in Figure 4.

Figure 5 shows that both Faster R-CNN and

YOLOv5 detect many regions occurring Campy-

lobacter bacteria, whereas RetinaNet fails to detect.

By comparing the results of Faster R-CNN with

those of YOLOv5, YOLOv5 detects smaller Cam-

Object Detection as Campylobacter Bacteria and Phagocytotic Activity of Leukocytes in Gram Stained Smears Images

537

Faster R-CNN

RetinaNet

YOLOv5

Figure 5: The result of detecting Camphylobacter bacteria

by Faster R-CNN, RetinaNet and YOLOv5.

phylobacter bacteria which Faster R-CNN cannot de-

tect. Also, Faster R-CNN has the case detecting that

non-Campylobacter bacteria are Campylobacter bac-

teria. Hence, YOLOv5 is the most appropriate de-

tector. On the other hand, the value of AP50 for

YOLOv5 is 43.6%, which is the reason that many en-

twined Campylobacter bacteria exist as Figure 4.

Table 3 illustrates the average running time of de-

tectors for detecting objects of Campylobacter bacte-

ria in one image.

Table 3: The average running time (msec) for detecting ob-

jects of Campylobacter bacteria in one image.

detector time

Faster R-CNN 164.0

RetinaNet 71.8

YOLOv5 36.8

Table 3 shows that YOLOv5 is the fastest in the

three detectors, about the half of the average running

time of RetinaNet and about the quarter of that of

Faster R-CNN.

3.2 Detection of Phagocytotic Activity

of Leukocytes

For the detection of phagocytotic activity of leuko-

cytes, we use 91 training images including 10 valida-

tion images and 10 test images for total 101 images.

Then, Table 4 illustrates the values of mAP, AP50 and

AP75 to detect phagocytotic activity of leukocytes by

using three detectors.

Table 4: The values of mAP, AP50 and AP75 to detect

phagocytotic activity of leukocytes (%).

detector mAP AP50 AP75

Faster R-CNN 38.0 62.8 41.2

RetinaNet 18.4 30.0 18.8

YOLOv5 45.5 70.4 53.1

Table 4 shows that YOLOv5 has the largest values

of mAP, AP50 and AP75. In contrast to Table 2 in

Section 3.1, the values of AP75 is larger than those of

mAP in Table 4, the threshold δ such that the value of

APδ is equal to that of mAP is in the range of [75, 95].

Then, for the detected regions of phagocytotic activity

of leukocytes, the overlap between the predicted box

and the ground truth box is large as IoU.

Hence, from the viewpoint of the average pre-

cision, detecting objects of phagocytotic activity of

leukocytes is more successful than detecting objects

of Campylobacter bacteria.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

538

Figure 6: The Gram stained smear image such that phago-

cytotic activity are annotated correctly, where phagocytotic

and non-phagocytotic regions are enclosed by red and yel-

low colors, respectively.

Table 5: The values of AP50 for phagocytotic (pha), non-

phagocytotic (non) and quasi-phagocytotic (quasi) images

(%).

detector pha non quasi

Faster R-CNN 83.7 57.3 47.5

RetinaNet 32.1 28.1 29.7

YOLOv5 92.7 59.5 59.1

Next, we represent the results of detecting phago-

cytotic activity of leukocytes in Gram stained smear

images. Figure 6 illustrates the Gram stained smear

image such that phagocytotic activity are anno-

tated correctly, where the phagocytotic and the non-

phagocytotic regions are enclosed by red and yellow

colors, respectively.

Then, Figure 7 illustrates the result of detecting

phagocytotic activity of leukocytes by Faster R-CNN,

RetinaNet and YOLOv5 from the Gram stained smear

images in Figure 6. Here, the phagocytotic images are

labeled by “true,” the quasi-phagocytotic images by

“false” and the non-phagocytotic images by “no.”

Figure 7 shows that YOLOv5 is the most appro-

priate detector for phagocytoticactivity of leukocytes,

which detects almost leukocytes with correct classes.

Faster R-CNN detects almost leukocytes but leuko-

cytes with incorrect classes and non-leukocytes sub-

stances as leukocytes. On the other hand, RetinaNet

is insufﬁcient to detect leukocytes.

Table 5 illustrates the values of AP50 for phago-

cytotic, non-phagocytotic and quasi-phagocytotic im-

ages.

Table 5 shows that the value of AP50 for phagocy-

totic images by YOLOv5 is much larger than those by

Faster R-CNN

RetinaNet

YOLOv5

Figure 7: The result of detecting phagocytotic activity of

leukocytes by Faster R-CNN, RetinaNet and YOLOv5.

Object Detection as Campylobacter Bacteria and Phagocytotic Activity of Leukocytes in Gram Stained Smears Images

539

Faster R-CNN and RetinaNet. Since the purpose of

this paper is to detect phagocytotic images correctly,

YOLOv5 is the most appropriate detector of the pur-

pose. Also, the values of AP50 for non-phagocytotic

images by all the detectors are very small. The rea-

son is that quasi-phagocytotic images is misclassiﬁed

to non-phagocytotic images. Nevertheless, if we re-

gard both quasi- and non-phagocytotic images as non-

phagocytotic images, this misclassiﬁcation can be ig-

nored to detect phagocytotic images correctly. As a

result, YOLOv5 succeeds to detect phagocytotic im-

ages correctly.

Table 6 illustrates the average running time of de-

tectors for detecting objects phagocytotic activity of

Leukocytes in one image.

Table 6: The average running time (msec) for detecting ob-

jects of phagocytotic activity of leukocytes in one image.

detector time

Faster R-CNN 146.0

RetinaNet 69.6

YOLOv5 34.9

Table 6 shows that YOLOv5 is the fastest in the

three detectors, about the half of the average run-

ning time of RetinaNet and about the quarter of that

of Faster R-CNN. By incorporating Table 3 in Sec-

tion 3.1 with Table 6, we can conclude that YOLOv5

is the fastest in the three detectors for not only

Campylobacter bacteria but also phagocytotic activ-

ity of Leukocytes.

Furthermore, the running time of detecting objects

of Camphylobacter bacteria is slightly larger than that

of phagocytotic activity of leukocytes but they are al-

most equal. Hence, the running time of the object

detection in this paper is independent from the object

as the target.

4 CONCLUSION AND FUTURE

WORKS

In this paper, we have detected Campylobacter bac-

teria and phagocytotic activity of leukocytes in Gram

stained smear images by using the detectors of Faster

R-CNN, RetinaNet and YOLOv5. Then, RetinaNet

have failed to detect them, and YOLOv5 is more ap-

propriate to detect them than Faster R-CNN.

In particular, for phagocytotic activity of leuko-

cytes, YOLOv5 have succeeded to detect almost

leukocytes with correct classes. On the other hand,

YOLOv5 have succeeded to detect Campylobac-

ter bacteria in many cases, whereas the cases that

Campylobacter bacteria have not detected exist. The

reason is that YOLOv5 does not work well for the im-

ages having many small objects.

Then, it is a future work to improve YOLOv5 to

work well for such images. In particular, we apply

YOLOv5 after decreasing the number of objects by

dividing an image and enlarging the objects.

Whereas YOLOv5 succeed to detect phagocytotic

images at 90% under AP50 in Table 5 as stated

in Section 3.2, there exist some images that non-

phagocytotic images are detected as phagocytotic.

Figure 8 illustrates such images.

Figure 8: The images that non-phagocytotic images are de-

tected as phagocytotic.

The upper-right region labeled by “no” in the left

image in Figure 8 and the lower-left and lower-right

regions labeled by “no” in the right image in Figure 8

are not leukocytes. Even if recall is more important

than precision in the medical data, it is a future work

to solve this misclassiﬁcation by improving annota-

tions in test images.

Since the number of training images in this paper

is too small to succeed object detection, it is neces-

sary to collect the large number of training images by

the medical technologist. Also since we use Gram

stained smears images photographed under the same

environment, it is necessary to collect the images un-

der the several environment, by using different equip-

ment and differentbrightness. These are future works.

REFERENCES

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

Bacteriol. Rev., 16:1–29.

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.

Everingham, M., Gool, L. V., Williams, C. K. I., Winn, J.,

and Zisserman, A. (2010). The PASCAL visual object

classes (VOC) challenge. Internat. J. Comput. Vision,

88:303–338.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

540

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

Yokoyama, S. (2020). Detecting Geckler classiﬁca-

tion from Gram stained smears images for sputum. In

Proc. ICPRAM’20, pages 469–476.

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

Yokoyama, S. (2020). Detection system of Gram

types for bacteria from Gram stained smears images.

In Proc. ICPRAM’20, pages 477–484.

Jocher, G. (2020). YOLOv5. https://github.com/ultralytics/

yolov5.

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

segmentation, detection and classiﬁcation of Gram

stained bacteria in blood sample. Master Thesis,

Lund University.

Lin, T., Goyal, P., Girshick, R., He, K., and Dollar, P.

(2017). Focal loss for dense object detection. In Proc.

ICCV’17, pages 2999–3007.

Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster R-

CNN: Towards real-time object detection with region

proposal networks. In Proc. NIPS’15, pages 91–99.

Simonyan, K. and Zisserman, A. (2015). Very deep convo-

lution networks for large-scale image recognition. In

Proc. ICLR’15.

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.

Wu, Y., Kirillov, A., Massa, F., Lo, W.-Y., and Gir-

shick, R. (2019). Detectron2. https://github.com/

facebookresearch/detectron2/.

Yoshihara, K. and Hirata, K. (2021). Detecting Camphy-

lobacter bacteria and phagocytotic activity of leuko-

cytes from Gram stained smear images. In Proc.

ESKM’21, pages 10–15.

Object Detection as Campylobacter Bacteria and Phagocytotic Activity of Leukocytes in Gram Stained Smears Images

541